Infrared image super-resolution reconstruction method based on noise decoupling

Abstract

The present application relates to the technical field of super-resolution reconstruction, and provides an infrared image super-resolution reconstruction method based on noise decoupling, comprising the following steps: S1: input data and pretreatment: using a FLIR infrared dataset, down-sampling a high-resolution image to a low-resolution input of 64*64 or 32*32 through bicubic interpolation, and normalizing pixel values to [0, 1] to adapt to a low-resolution infrared image I LR as subsequent input, and performing noise decoupling on the same. LR The present application has significant advantages in the task of infrared image super-resolution reconstruction, can effectively restore the details of an infrared image, and improve the resolution and quality of a reconstructed image; is suitable for the task of infrared image super-resolution reconstruction of different imaging devices and noise levels, and has important application value in fields such as medical imaging and environmental detection which have high requirements for image resolution.

Description

Technical Field

[0001] This invention belongs to the field of super-resolution reconstruction technology, specifically relating to a method for super-resolution reconstruction of infrared images based on noise decoupling. Background Technology

[0002] Infrared image super-resolution reconstruction has wide applications in remote sensing, nighttime surveillance, and medical imaging. However, due to limitations imposed by the physical characteristics of infrared sensors and environmental factors, high-resolution infrared data is often difficult to acquire. Furthermore, traditional super-resolution methods typically rely on large-scale paired high- and low-resolution data for supervised learning, while infrared image data acquisition is costly and annotation is difficult. Existing visible light-based methods (such as DDRM and DiffPIR) suffer from artifacts (such as stripe noise) and spectral aliasing when directly transferred to infrared images because they ignore the non-Gaussian nature of thermal noise and band response differences.

[0003] The Chinese patent publication number is "CN114913069A", entitled "A Super-Resolution Reconstruction Method for Infrared Images Based on Deep Neural Networks". This method effectively enhances the nonlinear fitting ability of the dataset and the network by using various data augmentation methods for low-resolution infrared images. In addition, multi-path learning is achieved in the residual blocks through parallel connections of convolutional kernels of different sizes, which improves the learning ability of the network and enables the network to learn local and global features simultaneously. However, the designed super-resolution network model is difficult to recover infrared image details, has poor generalization ability in various scenarios, and its performance drops significantly when there is no large-scale paired low-resolution and high-resolution dataset. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the aforementioned problems in existing technologies, this invention provides an infrared image super-resolution reconstruction method based on noise decoupling. This method solves the problems of existing infrared image super-resolution methods requiring specialized training for specific devices or noise conditions, having limited generalization ability, resulting in poor recovery performance and insufficient robustness under different sensors or high-noise environments.

[0006] (II) Technical Solution

[0007] The technical solution of this invention to solve the above problems is to provide a method for super-resolution reconstruction of infrared images based on noise decoupling, comprising the following steps:

[0008] S1: Input Data and Preprocessing: Using the FLIR infrared dataset, high-resolution images are downsampled to a low-resolution input of 64×64 or 32×32 using bicubic interpolation, and pixel values ​​are normalized to [0,1] to adapt to the low-resolution infrared image I of the neural network.LR As subsequent input, and for this I LR Perform noise decoupling;

[0009] S2: Pre-trained model loading: Load the pre-trained consistency model as the denoising network. Through feature extraction, residual block grouping, noise conditional embedding, and back projection guidance modules, calculate the noise mean μ of the input noise. n and standard deviation σ n Multilayer perceptron (MLP) is used to analyze normalized noise statistics. Encoding generates noise conditions This is then added channel by channel to the feature map, and finally the infrared high-frequency feature component F is output through back projection iteratively. HR Used as subsequent conditional input, the bottom feature extraction layer is frozen when the model is loaded, and only the high-level residual block and noise embedding layer are fine-tuned to adapt to the brightness distribution and noise characteristics of the infrared image;

[0010] S3: Constructing a Super-Resolution Model: Based on a multi-scale feature extraction module, a heterogeneous high-frequency information fusion module, and sub-pixel convolution, a super-resolution model is constructed to enhance the detail recovery capability of infrared images; low-resolution image I LR Super-resolution features F are generated by the multi-scale feature extraction module. SR And the infrared high-frequency features F generated by the consistency model denoising HR Cross-domain fusion is performed, and finally, a super-resolution image I is generated through sub-pixel convolution. SR ;

[0011] S4: Model Training: The low-resolution image obtained in S1 is decoupled from noise to generate initial noise N. LR As a conditional input, it is fed into the denoising network model in S2 for pre-training, and then back-projected to correct it in order to extract the infrared high-frequency feature vector F. HR Subsequently, the low-resolution image in S1 is decomposed using a multi-scale feature extraction module to obtain the super-resolution feature component F. SR ; For F HR and F SR Heterogeneous feature information fusion and sub-pixel convolution are performed to finally obtain the super-resolution reconstructed image I. SR ;

[0012] S5: Model Optimization and Loss Function Design: Composite loss functions include the mean squared error loss function (L... MSE ), perceptual loss function (L) LPIPS Consistency loss function (L) Con ) and gradient regularization loss function (L Reg We will perform joint optimization to ensure that the final output super-resolution infrared image achieves optimal quality.

[0013] Furthermore, in S1, the low-resolution infrared image I LR After three levels of Discrete Wavelet Transform (DWT), the basis function is Symlet4, decomposing it into a low-frequency subband {LL3} and a high-frequency subband; then, an adaptive hard thresholding function is used. High-frequency subband noise components are extracted; subsequently, the noise component z is reconstructed using inverse wavelet transform (IDWT). LR This achieves noise decoupling.

[0014] Furthermore, in S2, the low-resolution noisy image z LR The input denoising network includes a feature extraction layer, a residual block group, a noise conditional embedding layer, and a back projection guidance module. Feature extraction uses 3×3 convolutions combined with ReLU activation for initial feature extraction. The residual block group contains four residual units, each consisting of a 3×3 convolution, ReLU activation, and skip connections. Skip connections enhance the recovery of high-frequency details. The noise conditional embedding layer adjusts the noise level (1+δ)τ. n Encode the data and fuse it with the feature map to achieve adaptive denoising, specifically as follows:

[0015] The noise condition embedding: the noise level (1+δ)τ used in the denoising process is... n With the noise level τ injected into the image n Separate and control the noise independently, and split the noise into inverse correlation estimation noise. and random noise To achieve a fine-grained model of noise, specifically:

[0016]

[0017] Ensure that noise intervention preserves historical information while also being capable of handling random disturbances, balancing the restoration of details with the avoidance of local optima;

[0018] The back projection guidance module, based on data consistency correction, ensures that the reconstructed image possesses rich details while maintaining consistency with the degradation relationship of the input image by downsampling and matching the original observations. Specifically, this is expressed as follows:

[0019]

[0020] Furthermore, the multi-scale feature extraction module in S3 uses 3×3, 5×5, and 7×7 convolutional kernels to extract multi-scale features in parallel, and fuses information from different scales through Softmax normalized weights;

[0021] The heterogeneous high-frequency information fusion module performs linear transformation. Align the feature domains and compute fused features based on the channel attention mechanism. Finally, based on the fusion weight γ = Sigmoid(FC([F HR ,F SR ]))Calculate F final =γ·F fusion +(1-γ)·F SR .

[0022] Furthermore, the mean square error loss function (L) in S5 MSE ): Constrain super-resolution images to approximate real images at the pixel level, ensuring consistency in basic structure and brightness;

[0023] The perceptual loss function (L) LPIPS By performing similarity calculations in the feature space, the subjective perceptual quality of the reconstructed image is improved, and texture and higher-order semantic information are captured, making the reconstructed image superior in perceptual quality to the result of training using only MSE.

[0024] The consistency loss function (L) Con Consistent distillation ensures that the model's output remains stable at different time steps, reducing the impact of noise on super-resolution reconstruction.

[0025] The gradient regularization loss function (L) Reg ): Smooth the gradient during the optimization process to prevent gradient explosion or model overfitting;

[0026] The overall expression of the composite loss function is as follows:

[0027] L Total =λ1L MSE +λ2L LPIPS +λ3L Con +λ4L Reg

[0028] The loss weights are set as follows: λ1 = 1.0, λ2 = 0.1, λ3 = 0.5, and λ4 = 10. -4 .

[0029] (III) Beneficial Effects

[0030] Compared with existing technologies, this invention provides an infrared image super-resolution reconstruction method based on noise decoupling, which has the following beneficial effects:

[0031] 1. This invention proposes a noise decoupling and dynamic splitting injection strategy. By decoupling the noise component of the input low-resolution image from the image content and dynamically separating it into historical noise estimation terms and random perturbation terms, it solves the problem of detail blurring and artifacts caused by noise coupling in traditional super-resolution methods. This enables the model to make full use of noise statistical information and improve the stability and generalization ability of the denoising process.

[0032] 2. This invention provides a noise consistency correction strategy combined with back projection guidance. It corrects the noise output by the consistency model through back projection, calculates the error gradient of low-resolution noise, and guides the model to maintain statistical consistency with the input noise during denoising. This scheme solves the problem of noise distribution shift after model denoising, enabling the model to accurately retain the statistical characteristics of real noise and improve the realism and consistency of the denoised image.

[0033] 3. This invention solves the problems of difficult alignment of heterogeneous features and severe loss of details in infrared image super-resolution by end-to-end collaborative optimization of multi-scale feature extraction, noise decoupling guidance and high-frequency cross-domain fusion. This enables the model to adapt to different scenes and different noise intensities, and significantly improves the image quality and structural information integrity of super-resolution reconstruction. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart of the infrared image super-resolution reconstruction method based on noise decoupling described in this invention;

[0036] Figure 2 This is a framework diagram of the super-resolution model based on consistent denoising and multi-scale feature fusion of the present invention;

[0037] Figure 3 This is a structural block diagram of the consistency denoising network described in this invention;

[0038] Figure 4 This is a structural block diagram of the super-resolution structure of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Example

[0041] like Figure 1-4 As shown, an example of the present invention provides a method for infrared image super-resolution reconstruction based on noise decoupling, comprising the following steps:

[0042] S1: Input Data and Preprocessing: Prepare the FLIR infrared dataset for model testing, downsample the high-resolution images of the dataset using bicubic interpolation to generate a low-resolution input of 64×64 or 32×32, scale the pixel value range to [0,1] to fit the neural network input, and decouple the low-resolution degraded images from noise.

[0043] In the noise decoupling stage, the input low-resolution infrared image I... LR A three-level discrete wavelet transform (DWT) is used with Symlet4 as the basis function to decompose the wavelet into a low-frequency subband {LL3} and a high-frequency subband. First, noise components in the high-frequency sub-band are extracted using an adaptive threshold function. The hard threshold function is defined as follows:

[0044]

[0045] Wherein, the threshold of the kth layer σ k Let be the noise standard deviation of the k-th high-frequency sub-band, and N be the number of pixels in the sub-band. Then, the noise is reconstructed, retaining only the noise components of the high-frequency sub-band, and decoupled noise z is generated through inverse wavelet transform (IDWT). LR :

[0046] z LR =IDWT({0,Γ hard (LH1),Γ hard (HL1),Γ hard (HH1)},...,{0,Γ hard (LH3),Γ hard (HL3),Γ hard (HH3)}) Through the above decoupling strategy, the structural information of the infrared image is preserved, while the noise component is effectively separated, improving the noise modeling capability and providing higher quality input data for subsequent super-resolution reconstruction.

[0047] S2: Pre-trained model loading: Constructing a denoising network based on Consistency Models (CMs) θ A pre-trained consistency model is used as a denoising network to improve the stability and robustness of the infrared super-resolution reconstruction process.

[0048] The low-resolution noise z of the infrared image extracted after decoupling the noise in S1. LR The input is fed into a CMs denoising network for processing. The CMs network is as follows: Figure 3 As shown, a hierarchical denoising structure is adopted, including several key modules such as feature extraction, residual block grouping, noise conditional embedding, and reconstruction layer.

[0049] The feature extraction part uses 3×3 convolution combined with ReLU activation function for preliminary feature extraction to enhance the network’s ability to understand infrared noise. The feature extraction layer captures low-level features of infrared noise through convolution while preserving global context information, providing a foundation for subsequent noise modeling and high-resolution detail recovery.

[0050] In the noise-conditional embedding process, the mean noise value (μ) of the input is first calculated. n ) and standard deviation (σ) n ), Statistical modeling of input noise:

[0051] μ n ,σ n =f noise (z LR )

[0052] Subsequently, the residual block group consists of 4 residual units, each containing 3×3 convolution, ReLU activation and skip connections, to ensure effective information transfer and improve the high-frequency detail recovery capability of the image;

[0053] Noise conditional coding: Noise level (1+δ)τ is encoded using a fully connected layer. n Encode the data and extract the noise statistics (μ) from the infrared image. n ,σ n After normalization, the input is fed into a multilayer perceptron (MLP) to calculate the final noise conditions, adapting them to the infrared noise distribution under different scenarios.

[0054]

[0055] Calculated noise conditions As additional conditional information, it is added to the feature map channel by channel to achieve adaptive denoising for different noise intensities.

[0056] A noise-conditional injection strategy is adopted, and the injected noise is divided into inverse correlation estimation noise. With random noise z, ensure that noise intervention both preserves historical information and has the ability to handle random disturbances;

[0057] In each iterative optimization step, the current noise estimate is first calculated:

[0058]

[0059] in, It is the denoised noise of the current iteration state, z LR This represents the original noise term. This noise term reflects the update trend of the image in the current iteration state.

[0060] Then, independently sampled standard normal distribution noise. Then, the injected noise is generated by weighted combination using the hyperparameter η:

[0061]

[0062] Here, η is used to balance the influence of historical noise information and random noise. Through this mechanism, the present invention can effectively utilize existing noise estimation information while introducing an appropriate amount of random perturbation to prevent the model from getting trapped in local optima and improve the ability to restore image details. The feature map after injecting noise is used for the next step of back projection correction.

[0063] To ensure that the super-resolution image generated by the model maintains consistency with the original low-resolution observation data in terms of degradation relationship and to reduce information loss, this invention further designs a back-projection guidance module. First, degradation operator A is used to downsample the noise output of the consistency model to simulate the real image noise degradation process:

[0064]

[0065] Among them, z LR,est This represents the estimated noise after downscaling. Then, the estimated low-resolution noise z is calculated. LR,est With the low-resolution noise component z of the input LR The error between them is used to obtain gradient information:

[0066]

[0067] in, This is a pseudo-inverse operator used to correct the statistical consistency of noise generated by the model. In infrared image processing, noise is related to temperature radiation; therefore, further thermal radiation physical correction is applied to the noise to ensure the physical consistency of the infrared noise and its physical correlation with missing high-frequency components.

[0068]

[0069] Blackbody(·) is a pixel radiance correction based on Stefan-Boltzmann law.

[0070] In each iteration, the high-frequency information of the super-resolution is updated using this gradient information:

[0071]

[0072] Where, μ n The step size parameter is dynamically adjusted, and an adaptive step size adjustment strategy is adopted to optimize it based on the current gradient magnitude, τ. n-1 Controlling the intensity of noise disturbance, z comb The combined noise provided by the noise splitting module is used to prevent getting trapped in local optima.

[0073] After the above processing, the denoised infrared high-frequency feature vector F is output. HR This is to ensure the consistency of the image value range and avoid information loss and reconstruction distortion;

[0074] In addition, when loading the pre-trained model, the bottom feature extraction layer is frozen to preserve the global feature representation, and only the high-level residual block and noise embedding layer are fine-tuned to adapt to the brightness distribution and noise characteristics of the infrared image;

[0075] Through the above model design and pre-training process, the consistency model can efficiently extract structural information from infrared images, complete denoising within a few iterations, and adapt to different noise intensities and scenes, providing high-quality input for subsequent super-resolution reconstruction.

[0076] S3: Constructing a super-resolution model: such as Figure 4 As shown, the model consists of multiple functional modules, aiming to improve the detail recovery capability of low-resolution images through iterative optimization while ensuring data consistency. Specifically, the model includes: low-resolution image I... LR Super-resolution feature components F are generated by the multi-scale feature extraction module. SR The infrared high-frequency feature component F generated by the consistency model denoising in S2 HR Heterogeneous high-frequency component information is fused, and finally a super-resolution image I is generated through sub-pixel convolution. SR ;

[0077] Multi-scale feature extraction module: To effectively recover texture details at different scales in infrared images and address the difficulty of extracting cross-scale features using single-scale convolution, this invention introduces a multi-scale feature fusion method to enhance the ability to recover details at different scales. This module first uses 3×3, 5×5, and 7×7 convolution kernels in parallel to extract features from the original low-resolution image I. LR Parallel feature extraction is performed to generate multi-scale feature maps. These feature maps are then concatenated and weights are calculated using a fully connected layer. The weights are normalized using the Softmax function. The fused multi-scale features are then input into multiple residual blocks to further enhance image detail. Each residual block consists of a 3×3 convolution, a ReLU activation function, and skip connections, effectively extracting high-frequency details while preserving the integrity of feature propagation. The infrared feature F output from S2 is then removed. HR The enhanced feature F is obtained SR As super-resolution feature components, they enter the next heterogeneous information fusion module.

[0078] Heterogeneous high-frequency information fusion module: This module integrates the high-resolution features F from the consistent denoising output of S2. HR With multi-scale super-resolution features F SRTo achieve cross-domain integration. Due to F HR and F SR Due to heterogeneity, feature domain alignment is required between the two systems. A linear transformation is employed. Perform feature mapping to align the two elements in the same feature space:

[0079]

[0080] Calculate the statistical information of the transformed features, including the mean μ and variance σ:

[0081]

[0082] Subsequently, channel attention weights of the two feature maps are calculated based on the attention mechanism to enhance the ability to represent high-frequency information. The transformed features are then subjected to a one-dimensional convolution operation and normalized using the Softmax function to obtain the channel attention weights.

[0083]

[0084] The features are weighted and fused using the channel attention weights described above to obtain heterogeneous high-frequency weighted features:

[0085]

[0086] The final fusion weights are calculated using a channel attention mechanism to dynamically adjust the contribution ratio of infrared high-frequency features and super-resolution features:

[0087] γ=Sigmoid(FC([F HR ,F SR ]))

[0088] Finally, the final fusion features are calculated based on the fusion weights:

[0089] F final =γ·F fusion +(1-γ)·F SR

[0090] Through the above steps, deep fusion of infrared high-frequency information and super-resolution features is achieved, thereby enhancing the detail representation capability of the super-resolution reconstructed image. Finally, sub-pixel convolution is used for upsampling to generate the final super-resolution image I. SR :

[0091] I SR =SubpixelConv(F fusion )

[0092] Through the synergistic effect of the above modules, the image clarity is gradually improved during the iterative optimization process, while maintaining data consistency and structural integrity, thereby effectively solving the problems of detail loss and instability in traditional super-resolution methods.

[0093] S4: Model Training: The overall model training framework is as follows Figure 2 As shown, the low-resolution image obtained by S1 is decoupled from noise to generate initial noise N. LR As a conditional input, it is fed into the denoising network model in S2 for pre-training, and then back-projected to correct it in order to extract the infrared high-frequency feature vector F. HR Subsequently, the low-resolution image in S1 was decomposed using a multi-scale feature extraction module to obtain the super-resolution feature component F. SR For F HR and F SR Heterogeneous feature information fusion and sub-pixel convolution are performed to finally obtain the super-resolution reconstructed image I. SR .

[0094] The AdamW optimizer (β1 = 0.9, β2 = 0.999) was used during the optimization process, with an initial learning rate set to 2 × 10⁻⁶. -4 And a cosine annealing strategy is used for gradual decay:

[0095]

[0096] The minimum learning rate is set to 1×10. -5 The training cycle is 50 rounds, and the weight decay parameter is set to 10. -4 To prevent parameter overfitting, the model completes super-resolution reconstruction within no more than four neural network computations (NFE), with loss weight coefficients set to λ1 = 1.0, λ2 = 0.1, λ3 = 0.5, and λ4 = 10. -4 The weight coefficients were optimized through cross-validation and experiments to ensure the balance of the loss function, thereby improving both pixel-level reconstruction accuracy and perception quality. During training, hyperparameters were dynamically adjusted based on the peak signal-to-noise ratio (PSNR) evaluation metric. To ensure the stability of model training, an early stopping strategy was adopted to control training termination: if the PSNR on the validation set did not improve in 5 consecutive training rounds, training was terminated to ensure that the final output is a high-quality infrared super-resolution image.

[0097] S5: Model Optimization and Loss Function Design: To ensure high-quality infrared image super-resolution reconstruction in a few iterations, this method designs a weighted joint loss function, namely a composite loss function;

[0098] Using the mean squared error loss function (L MSEThis constrains the super-resolution image to approximate the real image at the pixel level, ensuring consistency in basic structure and brightness:

[0099]

[0100] Among them, I HR,i For true high-resolution images, I SR,i This represents the final output of the super-resolution network, where N is the number of training samples.

[0101] Introducing the perceptual loss function (L LPIPS ), L LPIPS The loss function better captures texture and higher-order semantic information, resulting in reconstructed images that outperform those trained using only MSE in terms of perceptual quality.

[0102]

[0103] Where VGG(·) represents the feature representation extracted by the pre-trained VGG network, and the loss calculates the Euclidean distance between the super-resolution image and the ground image in the VGG feature space.

[0104] Introducing the consistency loss function (L Con Consistent distillation ensures that the model's output remains stable across different time steps, reducing the impact of noise on super-resolution reconstruction.

[0105]

[0106] Among them, f θ For the consistency model, I t and The results are the denoising results at different time steps t and t′, respectively.

[0107] Introducing gradient regularization loss function (L Reg To avoid unstable optimization caused by excessively large gradient updates during training:

[0108]

[0109] in, This represents the gradient of the loss function with respect to the model parameters.

[0110] Considering all the loss functions mentioned above, the overall composite loss function is expressed as follows:

[0111] L Total =λ1L MSE +λ2L LPIPS +λ3L Con +λ4L Reg

[0112] The loss weights are set as follows: λ1 = 1.0, λ2 = 0.1, λ3 = 0.5, and λ4 = 10. -4 To ensure a balance between reconstruction quality and perception effect, this patent achieves efficient super-resolution reconstruction of infrared images through the above-mentioned loss function design and optimization strategies, balancing accuracy, speed, and robustness.

[0113] This invention achieves high-quality reconstruction of infrared images with minimal neural network computation through efficient infrared image super-resolution reconstruction based on noise decoupling, combined with back-projection guidance and optimized noise injection strategies. Compared to traditional super-resolution methods, this method offers advantages such as low computational cost, high reconstruction quality, and strong adaptability, making it suitable for infrared image applications in remote sensing, night vision, and medical imaging.

[0114] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for super-resolution reconstruction of infrared images based on noise decoupling, characterized in that, Includes the following steps: S1: Input Data and Preprocessing: Using the FLIR infrared dataset, high-resolution images are downsampled to a low-resolution input of 64×64 or 32×32 using bicubic interpolation, and pixel values ​​are normalized to [0,1] to adapt to the low-resolution infrared image I of the neural network. LR As subsequent input, and for this I LR Perform noise decoupling; S2: Pre-trained model loading: Load the pre-trained consistency model as the denoising network. Through feature extraction, residual block grouping, noise conditional embedding, and back projection guidance modules, calculate the noise mean μ of the input noise. n and standard deviation σ n Multilayer perceptron (MLP) is used to analyze normalized noise statistics. Encoding generates noise conditions This is then added channel by channel to the feature map, and finally the infrared high-frequency feature component F is output through back projection iteratively. HR Used as subsequent conditional input, the bottom feature extraction layer is frozen when the model is loaded, and only the high-level residual block and noise embedding layer are fine-tuned to adapt to the brightness distribution and noise characteristics of the infrared image; S3: Constructing a Super-Resolution Model: Based on a multi-scale feature extraction module, a heterogeneous high-frequency information fusion module, and sub-pixel convolution, a super-resolution model is constructed to enhance the detail recovery capability of infrared images; low-resolution image I LR Super-resolution features F are generated by the multi-scale feature extraction module. SR And the infrared high-frequency features F generated by the consistency model denoising HR Cross-domain fusion is performed, and finally, a super-resolution image I is generated through sub-pixel convolution. SR ; S4: Model Training: The low-resolution image obtained in S1 is decoupled from noise to generate initial noise N. LR As a conditional input, it is fed into the denoising network model in S2 for pre-training, and then back-projected to correct it in order to extract the infrared high-frequency feature vector F. HR Subsequently, the low-resolution image in S1 is decomposed using a multi-scale feature extraction module to obtain the super-resolution feature component F. SR ; For F HR and F SR Heterogeneous feature information fusion and sub-pixel convolution are performed to finally obtain the super-resolution reconstructed image I. SR ; S5: Model Optimization and Loss Function Design: Composite loss functions include the mean squared error loss function (L... MSE ), perceptual loss function (L) PIPS Consistency loss function (L) Con ) and gradient regularization loss function (L Reg We will perform joint optimization to ensure that the final output super-resolution infrared image achieves optimal quality.

2. The infrared image super-resolution reconstruction method based on noise decoupling according to claim 1, characterized in that: The low-resolution infrared image I in S1 LR After three levels of Discrete Wavelet Transform (DWT), the basis function is Symlet4, decomposing it into a low-frequency subband {LL3} and a high-frequency subband; then, an adaptive hard thresholding function is used. Extract high-frequency subband noise components; Subsequently, the noise component z is reconstructed using inverse wavelet transform (IDWT). LR This achieves noise decoupling.

3. The infrared image super-resolution reconstruction method based on noise decoupling according to claim 1, characterized in that: In step S2, the low-resolution noisy image z LR The input denoising network includes a feature extraction, residual block group, noise conditional embedding, and back projection guidance module. Feature extraction employs 3×3 convolution combined with ReLU activation for initial feature extraction. The residual block group contains four residual units, each consisting of a 3×3 convolution, ReLU activation, and skip connections. Skip connections enhance the recovery of high-frequency details. The noise-conditional embedding layer adjusts the noise level (1+δ)τ. n Encode the data and fuse it with the feature map to achieve adaptive denoising, specifically as follows: The noise condition embedding: the noise level (1+δ)τ used in the denoising process is... n With the noise level τ injected into the image n Separate and control the noise independently, and split the noise into inverse correlation estimation noise. and random noise To achieve a fine-grained model of noise, specifically: Ensure that noise intervention preserves historical information while also being capable of handling random disturbances, balancing the restoration of details with the avoidance of local optima; The back projection guidance module, based on data consistency correction, ensures that the reconstructed image possesses rich details while maintaining consistency with the degradation relationship of the input image by downsampling and matching the original observations. Specifically, this is expressed as follows:

4. The infrared image super-resolution reconstruction method based on noise decoupling according to claim 1, characterized in that: The multi-scale feature extraction module in S3 uses 3×3, 5×5, and 7×7 convolutional kernels to extract multi-scale features in parallel, and fuses information from different scales through Softmax normalized weights. The heterogeneous high-frequency information fusion module performs linear transformation. Align the feature domains and compute fused features based on the channel attention mechanism. Finally, based on the fusion weight γ = Sigmoid(FC([F HR ,F SR ]))Calculate F final =γ·F fusion +(1-γ)·F SR .

5. The infrared image super-resolution reconstruction method based on noise decoupling according to claim 1, characterized in that: The overall representation of the composite loss function in S5 is as follows: L Total =λ1L MSE +λ2L LPIPS +λ3L Con +λ4L Reg The loss weights are set as follows: λ1 = 1.0, λ2 = 0.1, λ3 = 0.5, and λ4 = 10. -4 .

Images