An infrared image enhancement method and system based on Mamba2

Abstract

The present application relates to the technical field of infrared image enhancement, in particular to an infrared image enhancement method and system based on Mamba2. The present application comprises: constructing a double-branch generator structure, wherein the main branch realizes super-resolution reconstruction of the infrared image based on a multilevel MambaRRDB module; the auxiliary branch extracts the edge gradient information of the input image and generates a high-resolution gradient image through shallow reconstruction. The MambaRRDB module containing state space modeling is adopted as the backbone network for simultaneously modeling the local details and global structural features in the image, thereby improving the image reconstruction quality; the prior dehazing and Sobel operator are introduced in the auxiliary branch to highlight the edge information of small weak targets, and the edge information is fused with the features of the main branch in the reconstruction process to further enhance the target area saliency. The generator and the discriminator are constructed through the adversarial training strategy to optimize the realism and structure fidelity of the image reconstruction, and finally a high-quality infrared image fused with edge information is output.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image technology, and in particular to an infrared image enhancement method and system based on Mamba2. Background Technology

[0002] Infrared images are widely used in scenarios such as nighttime surveillance, military reconnaissance, drone navigation, and border defense early warning due to their advantages such as strong penetration, low dependence on lighting, and good target concealment. However, compared with visible light images, infrared images usually suffer from low resolution, sparse texture information, and low signal-to-noise ratio. Especially under complex background interference, small targets in the image are often submerged in the background, which seriously restricts the performance of subsequent detection, recognition, and tracking algorithms.

[0003] To improve the visual quality and target saliency of infrared images, traditional image enhancement methods mainly include histogram equalization, Retinex enhancement, and multi-scale CLAHE, which can enhance image contrast and edge information to some extent. However, these methods generally rely on manual parameter settings, lack adaptive capabilities, and often introduce additional noise when dealing with non-uniform illumination or low signal-to-noise ratio backgrounds, leading to blurred target boundaries or false background enhancement.

[0004] In recent years, with the rise of deep learning, image enhancement models based on convolutional neural networks (CNNs) have made significant progress in the field of visible light images. Models such as SRCNN, EDSR, and ESRGAN have been widely used for super-resolution reconstruction and detail restoration tasks. Although some methods have been attempted to be applied to infrared image enhancement, traditional CNN structures still have limitations in feature extraction capabilities and robustness due to the unique characteristics of infrared images, such as the lack of high-frequency textures, small target areas, and weak background features. Especially when the background of an infrared image is complex or the target contrast is extremely low, the enhancement results often fail to highlight the target area and may even exacerbate noise interference.

[0005] Furthermore, while existing augmentation methods based on Generative Adversarial Networks (GANs) possess a certain degree of adaptive augmentation capability, most methods fail to design network structures specifically for the small target characteristics of infrared images, making it difficult to simultaneously achieve both global modeling capabilities and detailed representation of local targets. Meanwhile, while models based on self-attention mechanisms, such as Transformer, exhibit excellent global perception capabilities, their high computational complexity and resource consumption limit their practical deployment in embedded infrared applications.

[0006] Therefore, there is an urgent need to design an infrared image enhancement method that integrates state-space modeling capabilities, has a lightweight structure, and can simultaneously enhance the overall image quality and small target edge features, so as to improve the applicability of infrared images in real complex environments and the detectability of targets. Summary of the Invention

[0007] This invention addresses the problem of low signal-to-noise ratio and blurred details in infrared remote sensing images, making them difficult to detect, by proposing an infrared image enhancement method based on an improved Mamba2 model. This method employs a dual-branch Generative Adversarial Network (GAN) architecture: the main branch performs super-resolution reconstruction of the original low-quality infrared image, while the auxiliary branch extracts and enhances the gradient detail information of the image. The two branches interact at the feature level and are fused at the output stage to generate a high-quality enhanced infrared image. Through this design, this invention aims to improve the clarity and visibility of small targets in infrared images, increase the contrast between the target and the background, and thus improve the performance of subsequent infrared small target detection.

[0008] To achieve the above objectives, the infrared image enhancement method provided by the present invention includes the following steps:

[0009] The input infrared image I is fed into a dual-branch generator structure. The main branch performs super-resolution reconstruction of the infrared image based on a multi-level MambaRRDB module. The auxiliary branch extracts edge gradient information from the input image and generates a high-resolution gradient map through shallow reconstruction. Specifically, the auxiliary branch first applies a dark channel dehazing algorithm to the infrared image I to obtain a clear image J. Then, the dehazed image J is input into the Sobel convolution operator to calculate the first derivatives in the horizontal and vertical directions and obtain the gradient magnitude map. The gradient magnitude map is upsampled and fused with the intermediate features of the main branch in terms of channel dimension. It is then fed into the residual module for edge information enhancement. Finally, it is fused with the output image of the main branch to output the final enhanced image. ;

[0010] Enhanced image The output is fed into the discriminator, which uses the GAN adversarial training mechanism to determine the quality difference between the enhanced image and the original image, and updates the generator parameters so that the enhanced image and the high-quality infrared image cannot be distinguished at the discriminator level.

[0011] Finally, the infrared image to be enhanced is input into the trained generator, which then outputs the enhanced high-resolution image.

[0012] Furthermore, the MambaRRDB module includes: multiple cascaded CAMamba state modeling units, a residual connection structure, and a set of scaling factors that control the residual strength. The output Y is represented as:

[0013]

[0014] in, The intermediate output extracted for stacked CAMamba state modeling units, where X is the input tensor. It is an adjustable hyperparameter between 0 and 1.

[0015] Furthermore, the CAMamba state modeling unit includes a Mamba2 two-dimensional state space modeling unit and a channel attention module. The state update method of the Mamba2 two-dimensional state space modeling unit is as follows:

[0016]

[0017] in, This represents the state value of pixel (i,j) in the k-th direction, including the horizontal, vertical, or diagonal directions. As input features, , , , Update the parameter matrix for the state.

[0018] Furthermore, the calculation method for obtaining a clear image J is as follows:

[0019]

[0020] in, This represents the pixel coordinates in the image, and A is the estimated atmospheric light intensity. Transmittance, This is the preset minimum transmittance threshold.

[0021] Furthermore, the gradient magnitude plot The calculation formula is as follows:

[0022]

[0023] Where J represents a clear image, and These represent the gradient components in the horizontal and vertical directions of the image, respectively.

[0024] Furthermore, the gradient map G obtained from the auxiliary branch is enlarged to the same resolution and size as the feature map F of a certain layer in the main branch using bicubic interpolation. Then, using... Convolution compresses the number of channels in G to be equal to that in F, and extracts the most important high-frequency components. Let's assume that after... The gradient features after convolution and nonlinear activation are represented as H. H and F are then concatenated along the channel dimension to form a fused feature [H, F]. Next, an application is performed on the concatenated feature. Convolution and ReLU activation are performed to obtain fused detailed features. Finally, the original gradient map G is added to the fused features through residual connections to achieve identity transfer of gradient information, resulting in the output feature out, which can be expressed by the formula:

[0025] .

[0026] Furthermore, the final output image of the main branch Enhanced graph with auxiliary branch output The final output image is generated after pixel-level fusion. ,Right now:

[0027]

[0028] in It is a constant used to adjust the weights of the main branch and the edge enhancement branch (i.e., the auxiliary branch) in the fused image.

[0029] Furthermore, in the main branch, a convolutional block is set before the MambaRRDB module to increase the dimensionality of the RGB3 channels of the input image, and a convolutional block is set after the MambaRRDB module to reduce the dimensionality of the high-dimensional features.

[0030] Furthermore, in the auxiliary branch, upsampling is achieved through convolutional layers. After passing through the residual module, edge information enhancement is achieved through convolutional layers and contrast-limited adaptive histogram equalization.

[0031] The present invention also provides an infrared image enhancement system based on Mamba2 state modeling, comprising:

[0032] The system includes a processor and a memory. The memory stores program instructions, and the processor calls the stored instructions in the memory to execute an infrared image enhancement method based on Mamba2 state modeling as described in the above technical solution.

[0033] This invention, through the aforementioned fusion operation, integrates high-frequency edge information from the gradient domain and low-frequency semantic information from the image domain into the feature map output by the auxiliary branch. Compared to directly using the original gradient map, this provides a richer and more robust representation of details. The fused gradient-enhanced features work together with the reconstruction result of the main branch in the final stage of image generation to produce the final high-resolution enhanced infrared image. In this way, the main branch provides global reconstruction and super-resolution information, while the auxiliary branch provides local edge details. The complementary advantages of both result in a final output infrared image that is both clear and faithful to the original, while also highlighting the outlines of small targets.

[0034] The generative adversarial network of this invention also includes a discriminator module. The discriminator employs a convolutional neural network structure, performing layer-by-layer convolutional downsampling on the input image to extract discriminative features, and outputting the probability of determining whether the image is a real image or an augmented generated image. The introduction of the discriminator allows the backbone augmentation network to learn more realistic image reconstructions through adversarial training. Specifically, during training, an alternating optimization adversarial training mechanism is used: on the one hand, the discriminator is fixed while the generator is updated to minimize the reconstruction error and adversarial error of the generated image; on the other hand, the generator is fixed while the discriminator is updated to accurately distinguish between real and fake images.

[0035] In summary, the Mamba2-based infrared image enhancement method provided by this invention, through its innovative dual-branch GAN architecture design, combines super-resolution reconstruction with gradient detail enhancement, effectively improving the clarity of small targets and the contrast between targets and background in infrared images. Compared with existing technologies, this invention has significant advantages: First, by utilizing the improved MambaRRDB backbone network to recover image details, it can still reconstruct high-quality images under low signal-to-noise ratio conditions; second, by introducing a gradient-sensitive enhancement branch to highlight target edges, it strengthens key target features without introducing excessive computational overhead; third, by combining an adversarial training mechanism, the generated enhanced images show improvements in both subjective visual quality and objective metrics, and the training process converges stably without gradient explosion or mode collapse. Therefore, this invention improves the visual quality of infrared images while providing more reliable, high-quality input for subsequent small target detection, possessing significant practical value and widespread applicability. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the overall network framework of the infrared image enhancement method of the present invention, showing the structure of the main branch and gradient enhancement branch of the generator and their fusion relationship.

[0037] Figure 2 This is a schematic diagram of the MambaRRDB module structure in the backbone enhancement network of this invention, showing the specific structure of feature extraction and reconstruction performed by embedding Mamba units within the residual dense block.

[0038] Figure 3 This is a schematic diagram showing the effect of dehazing preprocessing on an infrared image, illustrating the background noise suppression and contrast enhancement of the original infrared image under the action of the dehazing module.

[0039] Figure 4 The image shows the gradient magnitude of the infrared image after extraction by the Sobel operator, illustrating the representation of image edges in the gradient domain, with the brightness reflecting the magnitude of the gradient.

[0040] Figure 5 This is a structural diagram of the discriminator model used in this invention.

[0041] Figure 6 This is a flowchart of an embodiment of the present invention.

[0042] Figure 7 This is a schematic diagram of the enhancement results of the method of the present invention in a typical infrared scene. The comparison shows the visualization effect of weak targets before and after enhancement. (a), (b), and (c) are images before enhancement, (d), (e), and (f) are images after enhancement, and (g), (h), and (i) are three-dimensional heat maps. Detailed Implementation

[0043] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the features of the following embodiments can be combined with each other to form new implementation methods.

[0044] like Figure 6 As shown in the figure, an infrared image enhancement method based on Mamba2 provided by this invention specifically includes: inputting an infrared image I and feeding it into a dual-branch generator structure, wherein the main branch realizes super-resolution reconstruction of the infrared image based on a multi-level MambaRRDB module, and the auxiliary branch extracts the edge gradient information of the input image and generates a high-resolution gradient map through shallow reconstruction. Specifically, the auxiliary branch first applies a dark channel dehazing algorithm to the infrared image I to obtain a clear image J, then inputs the dehazed image J into the Sobel convolution operator to calculate the first derivatives in the horizontal and vertical directions and obtain a gradient magnitude map, upsamples the gradient magnitude map, and fuses it with the intermediate features of the main branch in the channel dimension, sends it into the residual module for edge information enhancement, and finally fuses it with the output image of the main branch to output the final enhanced image. ;

[0045] Enhanced image The output is fed into the discriminator, which uses the GAN adversarial training mechanism to determine the quality difference between the enhanced image and the original image, and updates the generator parameters so that the enhanced image and the high-quality infrared image cannot be distinguished at the discriminator level.

[0046] Finally, the infrared image to be enhanced is input into the trained generator, which then outputs the enhanced high-resolution image.

[0047] The following section provides a detailed explanation of the overall process, main branch, auxiliary branch, fusion processing, and discriminator.

[0048] Overall process: such as Figure 1As shown, the infrared image enhancement method of this invention employs a generative adversarial network (GAN) framework, comprising two main parts: a generator and a discriminator. The generator consists of a two-branch structure: a main branch for super-resolution reconstruction and an auxiliary branch for gradient detail enhancement. First, the input low-resolution infrared image is standardized in size and fed into both branches: one branch enters the main branch to enhance resolution and reconstruct details; the convolutional block before the MambaRRDB module upscales the RGB3 channels of the input image, while the convolutional block after the MambaRRDB module downscales the high-dimensional features; the other branch enters the auxiliary branch to extract the gradient features of the image. The two branches maintain information exchange during feature extraction and are fused at the end of the generator to obtain the enhanced high-resolution infrared image output. The discriminator receives the real infrared image and the enhanced image output by the generator as input, and uses a discriminant network to determine whether it comes from real data or generated data, guiding the generator's training. The entire method can be summarized as follows: the original infrared image is input, processed in parallel by the main network and the auxiliary branch, and a higher-quality infrared image is output under the adversarial constraints of the discriminator. Residual Block is a standard residual module, and CLAHE is a contrast-limited adaptive histogram equalization, which is a method to enhance image contrast.

[0049] Main branch MambaRRDB: Figure 2 The main branch of the generator, the MambaRRDB module, is shown in the image above. The top image shows the internal structure of the MambaRRDB module, and the bottom image shows the internal details of a single CAMamba module. This module borrows the basic idea of ​​RRDB in ESRGAN, which uses a "residual within a residual" structure to stack multiple basic units to fully extract features. The difference is that this method introduces CAMamba blocks as the main feature extraction units, replacing the dense residual blocks (RDB) inside the traditional RRDB. Internally, the CAMamba module consists of two parts: the left side is a global dependency extraction part based on Mamba2 sequence modeling, processed on a token-sequence-by-total-sequence basis, specifically composed of a state-space modeler and a reverse residual feedforward network; the right side is a local feature enhancement branch combining channel attention mechanisms, performing detail extraction and robustness enhancement within the local neighborhood. In the CAMamba module, an explicit DropPath is introduced. Within each module, a certain probability of unit deactivation is applied after the sequence modeling unit and convolutional unit, randomly discarding some channel activation values ​​during training. This is equivalent to regularizing the network, preventing the accumulation of activation functions across all layers from causing abnormal gradient amplification. Meanwhile, the method increases the weights of the residual connections within CAMamba, thereby mitigating feature degradation when multiple layers are stacked.

[0050] The MambaRRDB module includes: multiple cascaded CAMamba state modeling units, a residual connection structure, and a set of scaling factors that control the residual strength. The output Y is represented as:

[0051]

[0052] in, The intermediate output extracted for stacked CAMamba state modeling units, where X is the input tensor. is an adjustable hyperparameter between 0 and 1. Where convolution... The purpose of this is to perform feature fusion and channel compression. The preceding residual connections will increase the number of channels, and the final convolutional layer will compress the dilated number of channels back to the input number of channels to ensure dimension matching.

[0053] The CAMamba state modeling unit includes the Mamba2 two-dimensional state space modeling unit and the channel attention module. The state update method of the Mamba2 two-dimensional state space modeling unit is as follows:

[0054]

[0055] in, This represents the state value of pixel (i,j) in the k-th direction, including the horizontal, vertical, or diagonal directions. As input features, , , , Update the parameter matrix for the state.

[0056] Gradient-enhancing auxiliary branches: such as Figure 1 As shown, the gradient enhancement auxiliary branch works in parallel with the main branch to highlight target edges and details in infrared images. First, addressing the issue of infrared images being susceptible to atmospheric noise and haze, the auxiliary branch introduces a dehazing preprocessing module to preprocess the original infrared image, such as... Figure 3 As shown. The dehazing module can employ existing image enhancement techniques such as dark channel prior dehazing algorithms and histogram equalization to perform smoke removal and contrast stretching operations on the input image. In this embodiment, an improved dark channel prior dehazing algorithm is preferably used to obtain an estimate of the atmospheric transmittance of the image and restore the scene radiance. After processing by this module, the original hazy infrared image background becomes cleaner, the overall contrast distribution is improved, and a clearer basic input is provided for subsequent gradient extraction. Next, as... Figure 4 As shown, the auxiliary branch applies the Sobel operator to the dehazed image to extract the gradient magnitude map G. The Sobel operator convolves the image in both the horizontal and vertical directions to calculate the gradient components. and Then, the gradient magnitude of each pixel is calculated according to the formula, thus forming the gradient magnitude map G. Figure 4 In the gradient map shown, bright areas represent pixels with large gradient magnitudes, typically corresponding to the edges of target objects, while dark areas represent pixels with small gradients, typically corresponding to flat backgrounds or noise points. It can be seen that after gradient extraction, small targets in the original image are highlighted in the gradient map due to their obvious grayscale changes; conversely, random noise in the background is effectively weakened due to its low gradient magnitude. Thus, the gradient enhancement branch extracts a feature map in the image domain that highlights the target and suppresses the background, preparing for further detail enhancement.

[0057] The calculation method for obtaining a clear image J is as follows:

[0058]

[0059] in, This represents the pixel coordinates in the image, and A is the estimated atmospheric light intensity. Transmittance, This is the preset minimum transmittance threshold.

[0060] Gradient magnitude plot The calculation formula is as follows:

[0061]

[0062] Where J represents a clear image, and These represent the gradient components in the horizontal and vertical directions of the image, respectively.

[0063] After obtaining the gradient magnitude map G, it is fused with the features of the intermediate layer of the main branch. This invention implements this function using a simple residual module. Since the main branch extracts multi-scale image features at different levels, to avoid repeatedly building deep convolutional structures in the auxiliary branches, this paper fuses the gradient information into the intermediate layer of the main branch. The feature map F of a certain layer in the middle of the main branch is selected and fused with the gradient map G. First, the gradient map G is enlarged to the same resolution and size as F using bicubic interpolation. Then, using... Convolution compresses the number of channels in G to be equal to that in F, and extracts the most important high-frequency components. Let's assume that after... The gradient features after convolution and nonlinear activation are represented as H. H is then concatenated with the backbone feature F along the channel dimension to form the fused feature [H, F]. Next, an application is performed on the concatenated feature. Convolution and ReLU activation are performed to obtain the fused detailed features. Finally, the original gradient map G is added to the fused features through residual connections to achieve identity transfer of gradient information, resulting in the output feature out, which can be expressed by the formula:

[0064] .

[0065] Through this module, the edge details extracted by the auxiliary branch are effectively injected into the semantic features of the main branch, forming enhanced features that contain both target edges and image textures.

[0066] In the generator's output phase, this invention will reconstruct the main branch's results. Gradient image of gradient branch output The images are then fused to obtain the final enhanced image. The specific integration method is to... As a guide, it is multiplied by a small weighting factor and then added to Above, sharpen edge information; Input after channel splicing Convolutional layers, ReLU layers and Convolutional layers learn to obtain the final output image after fusion. That is, the final output image of the main branch. Enhanced graph with auxiliary branch output The final output image is generated after pixel-level fusion. ,Right now:

[0067]

[0068] in It is a constant used to adjust the weights of the main branch and the edge enhancement branch in the fused image.

[0069] Regardless of the fusion strategy employed, the principle is to fully utilize the edge details provided by the gradient image to improve the clarity and contrast separation of small targets in the final image without compromising the overall quality of the reconstructed backbone image.

[0070] Discriminator: The discriminator network structure of this invention is as follows Figure 5 The diagram shows a four-layer convolutional neural network. Each layer extracts features through a convolutional kernel and performs stride downsampling, gradually compressing the input image into a smaller discriminative feature space. Finally, a fully connected layer outputs a probability value between 0 and 1. In this embodiment, the discriminator D contains three convolutional layers and a discriminator head module, employing a PatchGAN structure for local image patch discrimination. The discriminator aims to classify real infrared images as real and generated enhanced images as fake; conversely, the generator aims to "deceive" the discriminator as real as possible. The two work together in an adversarial game to improve the quality and realism of the generated results.

[0071] Ultimately, through the aforementioned dual-branch collaborative processing, the generator of this invention can generate high-quality infrared enhanced images with prominent targets, such as... Figure 7 As shown in the figure. Experiments show that introducing a gradient enhancement auxiliary branch does not compromise the training stability of the network. On the contrary, due to the focus on small target regions in the reconstruction loss, the method of this invention converges faster and achieves a lower loss value with fewer iterations. The entire adversarial training is iterated for several rounds under the common constraint of the loss function until there is no significant improvement in the generated image quality and discrimination accuracy on the validation set. After training, the enhancement model obtained by this invention can be used independently in the inference stage. That is, only the infrared image to be enhanced needs to be input into the trained generator network to output the enhanced high-resolution image without the need for a discriminator. Due to the end-to-end training and good adversarial strategy, the enhancement model of this invention has good generalization ability for different types of infrared images and is suitable for small target infrared imaging enhancement under various complex backgrounds.

[0072] In summary, this invention achieves synergistic optimization of global modeling and detail restoration capabilities in image enhancement by introducing a lightweight residual-dense module (MambaRRDB module) constructed from Mamba2 sequence units. Simultaneously, it focuses on enhancing the response features of small targets by incorporating edge gradient branches, effectively improving image quality under low signal-to-noise ratio conditions. Experimental results show that this invention significantly improves the visibility of small targets, image sharpness, and target boundary separation capabilities on multiple typical infrared remote sensing datasets, demonstrating good engineering practicality and deployment value.

[0073] On the other hand, embodiments of the present invention also provide an infrared image enhancement system based on Mamba2 state modeling, comprising:

[0074] The system includes a processor and a memory. The memory stores program instructions, and the processor calls the stored instructions in the memory to execute an infrared image enhancement method based on Mamba2 state modeling as described in the above technical solution.

[0075] The specific embodiments described herein are merely illustrative examples of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to replace them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. An infrared image enhancement method based on Mamba2 state modeling, characterized in that, include: The input infrared image I is fed into a dual-branch generator structure. The main branch performs super-resolution reconstruction of the infrared image based on a multi-level MambaRRDB module. The auxiliary branch extracts the edge gradient information of the input image and generates a high-resolution gradient map through shallow reconstruction. Specifically, the auxiliary branch first applies a dark channel dehazing algorithm to the infrared image I to obtain a clear image J. Then, the dehazed image J is input into the Sobel convolution operator to calculate the first derivatives in the horizontal and vertical directions and obtain the gradient magnitude map. The gradient magnitude map is upsampled and fused with the intermediate features of the main branch in terms of channel dimension. It is then fed into the residual module for edge information enhancement. Finally, it is fused with the output image of the main branch to output the final enhanced image. ; The MambaRRDB module includes: multiple cascaded CAMamba state modeling units, a residual connection structure, and a set of scaling factors that control the residual strength. The output Y is represented as: in, The intermediate output extracted for stacked CAMamba state modeling units, where X is the input tensor. These are adjustable hyperparameters between 0 and 1; The CAMamba state modeling unit includes the Mamba2 two-dimensional state space modeling unit and the channel attention module. The state update method of the Mamba2 two-dimensional state space modeling unit is as follows: in, This represents the state value of pixel (i,j) in the k-th direction, including the horizontal, vertical, or diagonal directions. For input features, , , , Update the parameter matrix for the state; Enhanced image The output is fed into the discriminator, which uses the GAN adversarial training mechanism to determine the quality difference between the enhanced image and the original image, and updates the generator parameters so that the enhanced image and the high-quality infrared image cannot be distinguished at the discriminator level. Finally, the infrared image to be enhanced is input into the trained generator, which then outputs the enhanced high-resolution image.

2. The method according to claim 1, characterized in that: The calculation method for obtaining a clear image J is as follows: in, This represents the pixel coordinates in the image, and A is the estimated atmospheric light intensity. Transmittance, This is the preset minimum transmittance threshold.

3. The method according to claim 1, characterized in that: Gradient magnitude plot The calculation formula is as follows: Where J represents a clear image, and These represent the gradient components in the horizontal and vertical directions of the image, respectively.

4. The method according to claim 1, characterized in that: The gradient map G obtained from the auxiliary branch is enlarged to the same resolution and size as the feature map F of a certain layer in the middle of the main branch using bicubic interpolation. Then, using... Convolution compresses the number of channels in G to be equal to that in F, and extracts the most important high-frequency components. Let's assume that after... The gradient features after convolution and nonlinear activation are represented as H. H and F are then concatenated along the channel dimension to form a fused feature [H, F]. Next, an application is performed on the concatenated feature. Convolution and ReLU activation are performed to obtain fused detailed features. Finally, the original gradient map G is added to the fused features through residual connections to achieve identity transfer of gradient information, resulting in the output feature out, which can be expressed by the formula: 。 5. The method according to claim 1, characterized in that: The final output image of the main branch Enhanced graph with auxiliary branch output The final output image is generated after pixel-level fusion. ,Right now: in It is a constant used to adjust the weights of the main branch and the edge enhancement branch (i.e., the auxiliary branch) in the fused image.

6. The method according to claim 1, characterized in that: In the main branch, a convolutional block is set before the MambaRRDB module to increase the dimensionality of the RGB3 channels of the input image, and a convolutional block is set after the MambaRRDB module to reduce the dimensionality of the high-dimensional features.

7. The method according to claim 1, characterized in that: In the auxiliary branch, upsampling is achieved through convolutional layers. After passing through the residual module, edge information enhancement is achieved through convolutional layers and contrast-limited adaptive histogram equalization.

8. An infrared image enhancement system based on Mamba2 state modeling, characterized in that, include: The processor and memory, wherein the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the infrared image enhancement method based on Mamba2 state modeling as described in any one of claims 1-7.

Images