Artificial Intelligence-Based Infrared Image Denoising Method and System

Abstract

This invention discloses an infrared image denoising method and system based on artificial intelligence, belonging to the field of image denoising technology. The method includes: acquiring infrared image data, extracting frequency domain spatial features, performing Gaussian filtering and restoring the image space, determining a high-frequency feature map, extracting image background information to determine a low-frequency feature map, and concatenating the low-frequency feature map with the image background information using a generator. An encoder is then used to perform a two-dimensional transposed convolution operation based on a convolution transpose self-attention mechanism. The method described in this invention effectively preserves spatial structure information such as edges and textures by transforming the infrared image from the frequency domain to high-frequency and low-frequency feature decomposition. The high-frequency feature extraction based on a Gaussian high-pass filter can effectively preserve spatial structure information such as edges and textures. The attention mechanism can weight their importance and enhance the exploration of noise. For the image background, the attention module can identify areas with excessive brightness fluctuations, and the residual characteristics can synchronize the transfer of secondary features during the noise removal process.

Description

Technical Field

[0001] This invention relates to the field of image denoising technology, and in particular to an infrared image denoising method and system based on artificial intelligence. Background Technology

[0002] Infrared imaging technology, as an important detection method, has advantages over visible light images, such as better penetration and stronger anti-interference ability. It is especially suitable for target recognition and background suppression in complex lighting environments or night scenes. However, because infrared imaging equipment is easily affected by factors such as sensor noise, environmental thermal radiation fluctuations, and imaging optical systems during the detection process, the generated infrared images are often accompanied by significant noise pollution. Traditional infrared image denoising methods mainly rely on domain-knowledge-based mathematical models, such as Gaussian filtering, median filtering, and wavelet denoising. Although these methods can suppress noise to a certain extent, they often cause the loss of image details, especially edges and textures. With the recent application of deep learning technology to infrared image denoising, architectures such as convolutional neural networks (CNN) and generative adversarial networks (GAN) can learn the complex relationship between the input noisy image and the denoising result.

[0003] However, deep learning methods often neglect the comprehensive modeling ability of multi-scale features and task synergy in infrared images while denoising. Most infrared image processing techniques lack a full understanding of the independent role of high-frequency and low-frequency features in noise decomposition, resulting in significant coupling between noise and background brightness elements, ultimately leading to unsatisfactory denoising results. Secondly, there are still difficulties in the joint optimization of image brightness correction and denoising tasks. Existing processing schemes often cannot guarantee the global and local consistency of the generated image when optimizing multiple tasks simultaneously. In particular, the lack of dynamic weight allocation strategies for multi-scale feature fusion results in deficiencies in both detail and global balance in denoising performance. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an artificial intelligence-based infrared image denoising method and system to address the common problem that deep learning-based methods often neglect the comprehensive modeling capabilities of multi-scale features and task synergy in infrared images during denoising. Most infrared image processing techniques lack a sufficient understanding of the independent role of high-frequency and low-frequency features in noise decomposition, leading to significant coupling between noise and background brightness elements, ultimately resulting in unsatisfactory denoising effects. Furthermore, the joint optimization of image brightness correction and denoising tasks remains challenging. Existing processing schemes often fail to guarantee the global and local consistency of the generated image when optimizing multiple tasks simultaneously. In particular, the lack of dynamic weight allocation strategies for multi-scale feature fusion results in deficiencies in both detail and global denoising performance.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an infrared image denoising method based on artificial intelligence, comprising:

[0008] Infrared image data is acquired to extract frequency domain spatial features, which are then Gaussian filtered and restored back to the image space to determine the high-frequency feature map. Background information is extracted to determine the low-frequency feature map, which is then concatenated with the high-frequency feature map by a generator. An encoder based on the convolution transpose self-attention mechanism is used to perform a two-dimensional transpose convolution operation. Residual connections are used to obtain the jointly output denoised high-frequency feature map. A denoising filter is then used, and the brightness component is corrected to obtain the denoised and brightness-corrected image.

[0009] The loss is calculated on the image after denoising and brightness correction based on the supervised regularization mechanism. At the same time, the comprehensive loss is determined based on the physical consistency loss, the loss of the denoising task, and the loss of the brightness correction task. The generator parameters are iteratively optimized through gradient backpropagation to output a stable denoised and brightness corrected image.

[0010] The low-frequency background features, high-frequency detail features, and noise residual features are determined and the output feature map is spliced. An attention weight matrix is ​​generated and dynamically weighted according to the channel coordinate attention block algorithm. Global max pooling is used to extract global dynamic weights and generate an enhanced feature map. Dynamic weight allocation is performed based on the enhanced feature map and the weighted feature map. The low-frequency background features and high-frequency detail features are fused to obtain a dynamic fused feature map.

[0011] Multi-scale sampling is used to generate feature maps of different scales based on the dynamically fused feature maps. The dynamically fused feature maps are then re-inputted into the generator to generate an updated input image. The encoder and decoder are then used to fuse the acquired optimized denoised image and brightness-corrected image to obtain an optimized and enhanced image.

[0012] The optimized and enhanced image is iteratively re-output and analyzed by comparing the peak signal-to-noise ratio and structural similarity index with the corresponding threshold.

[0013] As a preferred embodiment of the artificial intelligence-based infrared image denoising method of the present invention, the method includes: extracting frequency domain spatial features from the acquired infrared image data, stitching them together using a generator, employing an encoder based on a convolutional transpose self-attention mechanism, combined with a denoising filter, and correcting the brightness component, including...

[0014] Infrared image data is acquired and converted to the frequency domain using a two-dimensional discrete Fourier transform. The amplitude is then calculated based on the amplitude and phase information of the frequency domain data at the corresponding frequencies.

[0015] A Gaussian high-pass filter was used as the filtering reference to determine the frequency domain image after high-frequency filtering, and a two-dimensional inverse Fourier transform was used to restore it back to the image space to obtain the high-frequency feature map.

[0016] A Gaussian low-pass filter is used to extract the background information of the image. A low-frequency filter is applied to the image in the frequency domain, and 2D-IDFT is used to restore it back to the image space to obtain the low-frequency feature map.

[0017] The extracted high-frequency and low-frequency images are verified against the original image and then stitched together to obtain a multi-channel input image.

[0018] The encoder is based on the Convolutional Transposed Self-Attention (CTASB) mechanism, where the encoder extracts features from the multi-channel input image through convolutional layers and performs feature weighting to obtain attention features. The decoder then performs a two-dimensional transposed convolution operation on the attention features to obtain the decoding output of different convolutional layers.

[0019] Using the residual connection method, some features of the encoder are directly fed back to the decoding stage to obtain a denoised high-frequency feature map of the joint output. Based on the denoised high-frequency feature map, a denoising filter is used for denoising processing, and a normalized activation function is used to correct the brightness component to obtain the denoised and brightness-corrected image.

[0020] As a preferred embodiment of the artificial intelligence-based infrared image denoising method of the present invention, the step of performing loss calculation on the denoised and brightness-corrected image, iteratively optimizing the generator parameters through gradient backpropagation, and outputting a stable denoised and brightness-corrected image includes:

[0021] Loss calculation is performed based on supervised regularization mechanism. Regularization terms are calculated for the first-order gradient of the denoised image and the second-order gradient of the corrected brightness image, respectively. The first-order gradient of the denoised image is constrained by L1 norm and the sum of gradient magnitudes is calculated. The second-order gradient of the corrected brightness image is constrained by L2 norm.

[0022] Based on the input image data, the physical consistency loss is calculated by combining the denoised image and the corrected brightness image. The comprehensive loss is determined by jointly optimizing the loss of the denoising task, the loss of the brightness correction task, the supervised regularization term, and the physical consistency loss.

[0023] The gradient is calculated through backpropagation based on the comprehensive loss, and gradient descent optimization is performed through the Adam optimizer to update the generator parameters. When the comprehensive loss no longer decreases significantly during continuous iteration, the iteration stops, the generator parameters are output, and a stable denoised image and a corrected brightness image are output.

[0024] As a preferred embodiment of the artificial intelligence-based infrared image denoising method of the present invention, wherein: the stitched output feature map is processed by using global max pooling to extract global dynamic weights, generating an enhanced feature map, and fusing low-frequency background features and high-frequency detail features to obtain a dynamically fused feature map, including...

[0025] Features are defined based on the denoised image and the corrected brightness image, including low-frequency background features, high-frequency detail features and noise residual features, and then the features are stitched together to output a feature map.

[0026] Based on the Channel Coordinate Attention Block (CCAB) algorithm, an attention weight matrix is ​​generated by calculating the statistical importance of the feature map in the channel dimension. The input feature map is dynamically weighted and adjusted using the attention weight matrix, and the weighted feature map is output.

[0027] Based on the feature map, global max pooling (GMP) is used to extract the global maximum value of each channel to generate an initial global feature matrix. This matrix is ​​then transformed through two global fully connected layers to generate a global dynamic weight matrix, which is used to generate an enhanced feature map.

[0028] Dynamic weight allocation is performed based on the enhanced feature map and the weighted feature map, and low-frequency background features and high-frequency detail features are fused to obtain a dynamic fused feature map.

[0029] As a preferred embodiment of the artificial intelligence-based infrared image denoising method of the present invention, the step of generating feature maps of different scales based on a dynamic fusion feature map using multi-scale sampling operations, and re-inputting the dynamic fusion feature map into the generator to generate an updated input image includes:

[0030] Based on the dynamically fused feature map, multi-scale sampling operations are used to generate feature maps of different scales. Based on the determined fusion weights, the feature maps generated at multiple scales are weighted and fused to obtain the final multi-scale features.

[0031] As a preferred embodiment of the artificial intelligence-based infrared image denoising method of the present invention, the step of reusing the encoder and decoder to perform data fusion on the acquired optimized denoised image and brightness-corrected image to obtain an optimized enhanced image includes:

[0032] The generator is re-inputted based on the dynamically fused feature map, and then stitched together with the high-frequency feature image and the low-frequency feature image to generate an updated input image.

[0033] Based on the updated input image, the encoder and decoder are reused to obtain the decoding output of different convolutional layers. A denoising filter is used to obtain an optimized denoised image and a corrected brightness image. Furthermore, data fusion is performed through the sigmoid function to obtain an optimized and enhanced image.

[0034] As a preferred embodiment of the artificial intelligence-based infrared image denoising method of the present invention, the iterative re-output of the optimized and enhanced image, based on the comparison and analysis of peak signal-to-noise ratio and structural similarity indices with corresponding thresholds, includes:

[0035] Based on the iterative re-output of the dynamically fused feature map, the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) of the optimized and enhanced image are calculated. Based on the reference real image in the validation set data, the signal-to-noise ratio threshold and the similarity threshold are determined. If the PSNR of the optimized and enhanced image is greater than the signal-to-noise ratio threshold and the SSIM of the optimized and enhanced image is greater than the similarity threshold, the iteration is stopped, and the optimized and enhanced image of the current iteration is used as the final processed image.

[0036] Secondly, the present invention provides an infrared image denoising system based on artificial intelligence, comprising,

[0037] The data preprocessing module acquires infrared image data and extracts information from different frequencies of the image to determine high-frequency feature maps and low-frequency background feature maps.

[0038] The generator module concatenates the initial multi-channel input image, extracts multi-scale semantic features, and gradually restores the image resolution using a two-dimensional transposed convolution operation. It also uses a denoising filter to further process and reduce noise, and corrects the brightness component.

[0039] The loss calculation and optimization module defines a comprehensive loss function and iteratively optimizes it through gradient backpropagation and Adam optimizer to finally output a stable denoised image and a brightness-corrected image.

[0040] The dynamic feature extraction module extracts low-frequency background features, high-frequency detail features, and noise residual features, splices them into a multi-channel feature map, generates an enhanced feature map through global max pooling, and generates a dynamic fused feature map.

[0041] The multi-scale sampling enhancement module generates feature maps of different scale ranges by dynamically fusing feature maps, and optimizes the denoised image and brightness correction image through encoder and decoder to generate an optimized and enhanced image.

[0042] The image quality assessment module calculates the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) of the optimized and enhanced images, and performs comparative analysis with the real images in the validation set.

[0043] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the artificial intelligence-based infrared image denoising method described in the first aspect of the present invention.

[0044] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the artificial intelligence-based infrared image denoising method described in the first aspect of the present invention.

[0045] The beneficial effects of this invention are as follows: By transforming the infrared image into high-frequency and low-frequency feature decomposition, high-frequency features extracted based on a Gaussian high-pass filter can effectively preserve spatial structure information such as edges and textures. The attention mechanism can weight their importance and enhance the exploration of noise. For the image background, the attention module can identify areas with excessive brightness fluctuations. The residual characteristics can synchronize the transmission of secondary features during the noise removal process, avoiding feature mutations or information loss during the decoding stage. Through the joint enhancement of the denoising filter and brightness correction, the denoising filter nonlinearly adjusts the intensity of high-frequency noise, suppressing residual noise while preserving high-frequency sharpening characteristics. The Gaussian parameter ensures that the noise processing intensity is controllable, avoiding excessive erasure of key details. The joint optimization of the comprehensive loss function improves task collaboration. The supervised regularization term strengthens the physical properties of the image from the perspective of gradient constraints. By extracting low-frequency information from the denoised image, low-frequency background features are formed to represent the smooth features and large-scale distribution of global illumination. Through dynamic weight allocation, low-frequency and high-frequency features are fused to ensure that the fusion result can achieve a balanced feature distribution in different scenarios. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the 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.

[0047] Figure 1This is a flowchart illustrating the artificial intelligence-based infrared image denoising method in Example 1.

[0048] Figure 2 This is a schematic diagram of the structure of the artificial intelligence-based infrared image denoising system in Example 1. Detailed Implementation

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0050] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0051] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0052] Example 1, referring to Figures 1 to 2 This is the first embodiment of the present invention, which provides an infrared image denoising method based on artificial intelligence, including the following steps:

[0053] S1: Acquire infrared image data, extract frequency domain spatial features, perform Gaussian filtering and restore back to image space to determine high-frequency feature map, extract image background information to determine low-frequency feature map, and stitch the generator with the high-frequency feature map. The encoder uses a convolution transpose self-attention mechanism to perform two-dimensional transpose convolution operation, and uses residual connection to obtain the joint output denoised high-frequency feature map. Combined with a denoising filter and correction of brightness components, the denoised and brightness corrected image is obtained.

[0054] Preferably, the infrared image data is acquired, frequency domain spatial features are extracted, and the data is stitched together using a generator. An encoder based on a convolutional transpose self-attention mechanism is employed, along with a denoising filter and correction of the brightness component.

[0055] Infrared image data is acquired and converted to the frequency domain using a two-dimensional discrete Fourier transform, as follows:

[0056]

[0057] Where F(u,v) represents the frequency domain coordinate data of the infrared image, M and N represent the number of rows and columns of the image, respectively, and I ip(x, y) represents the infrared image pixel data, u and v represent the frequency domain coordinates (horizontal and vertical frequencies) respectively, j represents the imaginary unit, and e -j2π(ux / M+vy / N) This represents the complex terms (phase information) of the corresponding frequency domain transformation, and their product completes the frequency weighting;

[0058] Based on the amplitude and phase information of the frequency domain spatial data at the corresponding frequency, the amplitude is calculated and expressed as:

[0059]

[0060] Where Re(F) and Im(F) represent the real and imaginary parts of the Fourier transform result, respectively, and |F(u,v)| represents the amplitude of the frequency domain spatial data;

[0061] Using a Gaussian high-pass filter as the filtering benchmark, the frequency domain image after high-frequency filtering is determined, and then a two-dimensional inverse Fourier transform is used to restore it back to the image space to obtain the high-frequency feature map, represented as:

[0062]

[0063] F hf (u, v) = H hf (u,v)·F(u,v);

[0064]

[0065] Where H hf (u,v) represents the weights of the Gaussian high-pass filter, used to determine the degree to which a certain frequency (u,v) is passed in the frequency domain. The closer the value is to 1, the more the frequency is preserved; the closer it is to 0, the more the frequency is attenuated. σ represents the Gaussian parameter that controls the attenuation amplitude of the filter, determined based on historical experience. F hf (u, v) represents the frequency components in the frequency domain after high-frequency filtering, I hf (x,y) represents the high-frequency feature map. This represents the inverse Fourier transform. The "1" in the formula represents the base value of the filter, and the low-frequency weights decrease starting from 1.

[0066] A Gaussian low-pass filter is used to extract background information from the image. A low-frequency filter is applied to the image in the frequency domain, and 2D-IDFT is used to reconstruct the image space, yielding a low-frequency feature map, represented as follows:

[0067]

[0068] F cf (u, v) = H cf (u,v)·F(u,v);

[0069]

[0070] Where H cf (u,v) represents the weights of the Gaussian low-pass filter, F cf (u, v) represents the frequency components in the frequency domain after low-frequency filtering, I cf (u,y) represents the low-frequency feature map;

[0071] The extracted high-frequency and low-frequency images are verified against the original image (ensuring that edge information is preserved for high-frequency components and background information is preserved for low-frequency components), and then stitched together to obtain a multi-channel input image, represented as follows:

[0072] I gen (x,y,c)=concat(I hf (x,y,c1),I hf (x,y,c2));

[0073] Where I gen (x,y,c) represents the multi-channel input image of the generator, with pixel coordinates (x,y) and channel label c. hf (x, y, c1) represents a high-frequency feature image with channel c1, I cf (x,y,c2) represents the low-frequency feature image, with channel c2;

[0074] The encoder is based on the Convolutional Transposed Self-Attention (CTASB) mechanism. The encoder extracts features from the multi-channel input image through convolutional layers and performs feature weighting to obtain attention features. The decoder then performs a two-dimensional transposed convolution operation on these attention features to obtain the decoding outputs of different convolutional layers, as shown below:

[0075]

[0076] in Let represent the output feature map of the i-th layer of the encoder, and Conv(·) denote a two-dimensional convolution operation used for preliminary feature extraction. Z represents the multi-channel input image features of the previous layer. cts (x, y, c) represents the attention features output by CTSAB, and Q, K, V represent the input features (query, key, value) mapped through the convolutional layer, respectively. T The transpose of the key is calculated, d k This represents the channel dimension of the feature map, used for normalizing attention summation, and is a fixed parameter in the encoder (taking the number of channels). This represents the output feature map of the i-th layer of the decoder. Among them Equal to Z cts(x,y,c),Deconv(·) represents the two-dimensional transpose convolution operation, used to increase the resolution layer by layer;

[0077] Using the residual connection method, some features from the encoder are directly fed back to the decoding stage to obtain the jointly output denoised high-frequency feature map, represented as:

[0078] I out (x,y)=I dec (x,y,c)+Z en (x,y,c);

[0079] Where I dec (x,y,c) and Z en (x, y, c) represent the output feature maps of the encoder and decoder, respectively. out (x,y) represents the high-frequency feature map of the noise reduction, which preserves the frequency domain decomposition features while effectively reducing noise;

[0080] Denoising is performed using a denoising filter based on the high-frequency feature map, and the brightness component is corrected using a normalized activation function. The resulting image after denoising and brightness correction is shown below:

[0081] D ce (x,y)=I out (x,y)·H den (x,y);

[0082]

[0083] D br (x,y)=I cf (x,y)+sigmoid(D ce (x,y));

[0084] Where H den (x,y) represents a filter adapted to the noise characteristics, D ce (x,y) represents the denoised image, τ represents the Gaussian parameter of the denoising filter attenuation magnitude, determined based on historical experience, and D br (x,y) represents the corrected brightness image.

[0085] By transforming the infrared image into a frequency domain and decomposing it into high-frequency and low-frequency features, high-frequency features can be extracted based on a Gaussian high-pass filter, which can effectively preserve spatial structural information such as edges and textures. The high-frequency part contains the main noise components of the image, but also carries edge features. This process allows the noise reduction task in the subsequent generator to focus on the high-frequency noise region without destroying the structural information of the image. By extracting low-frequency components through a Gaussian low-pass filter, the low-frequency image after wavelet extraction effectively removes high-frequency noise components, leaving only the illumination distribution and smooth background content. After removing high-frequency noise, the low-frequency feature map reduces interference and preserves the basic scene information.

[0086] By setting the encoder to feature weighting based on attention, the optimization efficiency is improved. The CTSAB module dynamically allocates the weights of spatial channels through the calculation of the attention matrix Q, K, V. This allows the attention mechanism to weight the importance of certain noise-concentrated areas in the image, such as high-frequency parts, and enhance the exploration of noise. For the background of the image, the attention module can identify areas with excessive brightness fluctuations and improve the distribution by weighting. This allows the generator to allocate more computing resources to key areas during denoising and brightness optimization, while reducing the computing overhead of non-important areas. In the decoding stage, the attention features of CTSAB are subjected to two-dimensional transposed convolution to increase the resolution layer by layer and restore the spatial structure of the image. The transposed convolution can reconstruct high-frequency features and restore the structural details of noise areas in depth. The weighted compensation of the attention features after CTSAB decoding further improves the image edge transition and background smoothness.

[0087] Residual connections enhance the stability and performance of the overall structure. They directly feed back some features from the encoder to the decoder, enabling the decoder to obtain additional compensation features when interpreting high-frequency features. High-frequency feature feedback from the encoder can prevent some noise removal tasks from excessively erasing edge details. Residual features can synchronize the transmission of secondary features during noise removal, preventing feature mutations or information loss during the decoding stage. When facing complex noise distributions, residual connections can effectively provide additional encoder support, enabling the generator to stably handle edge noise and high-frequency regions.

[0088] By combining the denoising filter and brightness correction, the denoising filter nonlinearly adjusts the intensity of high-frequency noise, suppressing residual noise while preserving high-frequency sharpening characteristics. The Gaussian parameter ensures that the noise processing intensity is controllable, avoiding excessive erasure of key details. The normalized activation function smooths the brightness distribution of the low-frequency background, making the background brightness more natural and the transition more uniform. This results in more thorough noise suppression in the final output image, while preserving edge details, making the background illumination more uniform, and the overall scene perception effect more natural.

[0089] S2 calculates the loss of the denoised and brightness-corrected image based on the supervised regularization mechanism. At the same time, it determines the comprehensive loss based on the physical consistency loss, the loss of the denoising task, and the loss of the brightness correction task. It iteratively optimizes the generator parameters through gradient backpropagation and outputs a stable denoised and brightness-corrected image.

[0090] Preferably, loss calculation is performed on the denoised and brightness-corrected image, and the generator parameters are iteratively optimized through gradient backpropagation to output a stable denoised and brightness-corrected image, including:

[0091] Loss calculation is performed based on supervised regularization. Regularization terms are calculated for the first-order gradient of the denoised image and the second-order gradient of the corrected brightness image, respectively. The first-order gradient of the denoised image is constrained by the L1 norm, and the sum of gradient magnitudes is calculated. The second-order gradient of the corrected brightness image is constrained by the L2 norm, as follows:

[0092]

[0093]

[0094] Where Ω represents the supervised regularization term. This represents the sum of gradient magnitudes. λ1 represents the L2 norm constraint applied to the second-order gradient of the corrected brightness image, λ2 represents the hyperparameter controlling the intensity of gradient regularization of the denoised image (value is set experimentally), and λ3 represents the hyperparameter controlling the intensity of second-order gradient regularization of the brightness recovery image (value is a fixed empirical value).

[0095] Based on the input image data, the physical consistency loss is calculated by combining the denoised image and the corrected brightness image, and is expressed as:

[0096] L ph =||D br (x,y)-I ip (x,y)+D ce (x,y)||1

[0097] Where L ph represents the physical consistency loss, used to constrain the consistency of imaging patterns between the generated image and the input image; ||·||1 represents the L1 norm, which is used to obtain the absolute error between the restored image and the input image.

[0098] The comprehensive loss is determined by jointly optimizing the loss from the denoising task, the brightness correction task, the supervised regularization term, and the physical consistency loss, and is expressed as:

[0099] L to =α1L de +α2L de +Ω+Lph ;

[0100] L de =||D ce (x,y)-I det ||1;

[0101] L re =||D br (x,y)-I iut ||1;

[0102] Where L to L represents the overall loss, where α1 and α2 are constants that balance the denoising and brightness correction tasks, respectively, determined based on historical experience. de L represents the loss of the denoising task. re I represents the loss in the brightness correction task. det I represents the denoised reference image. iut Indicates a brightness correction reference image;

[0103] The gradient is calculated through backpropagation based on the comprehensive loss, and gradient descent optimization is performed through the Adam optimizer to update the generator parameters (including encoder parameters, CTSAB module parameters, and decoder parameters). When the comprehensive loss no longer decreases significantly during continuous iteration, the iteration stops, the generator parameters are output, and a stable denoised image and a corrected brightness image are output.

[0104] By applying L1 norm constraints to the first-order gradient of the denoised image, i.e. calculating the sum of gradient magnitudes, the edge detail quality and smoothness of the generated image can be improved. At the same time, combined with L2 norm regularization of the second-order gradient of the brightness image, the discontinuity of brightness transition in the image is reduced, making the background area of ​​the generated image more uniform, while avoiding abrupt brightness jumps. The background brightness of the infrared image usually presents a smooth gradient rather than a drastic change. The second-order gradient constraint imposes a smoothness limit on the brightness correction result, making the brightness of the background area approximate the actual physical distribution law.

[0105] By combining the physical consistency loss with the input image, the combined result of the denoised image and the brightness restoration image is constrained with the original input image. From the perspective of imaging laws, the physical consistency loss ensures a one-to-one mapping between the generated image and the input image in terms of physical characteristics. By calculating the absolute error, it limits the generated image from deviating too much from the characteristics of the input image. In the denoising process, there may be loss of details due to filtering or noise residue due to insufficient filtering. The physical consistency loss directly compares the output and input of the entire image, providing a global balance constraint to ensure that the generator will not overfit to denoising or cause a global offset in brightness restoration. For infrared image applications (such as target detection or scene analysis), the consistency of imaging laws is an important measure of usability in real-world scenarios. The physical consistency loss limits the deviation of the generated image, making the denoised image more consistent with the real physical data.

[0106] By jointly optimizing the comprehensive loss function, task synergy is improved. The supervised regularization term strengthens the physical properties (edge ​​and smoothness) of the image from the perspective of gradient constraints, while clarifying the synergistic requirements of denoising and brightness correction in terms of physical properties. This supplementary constraint enables the generator to achieve optimization in local regions (such as noisy regions) and global regions (such as background illumination) without interfering with each other. Compared with the denoising loss and brightness correction loss, which only focus on the optimization of a certain branch, the physical consistency loss provides a balanced constraint from the perspective of the global output, ensuring the overall output effect of the model. Through the joint action with gradient regularization, the generator explicitly learns more physically reasonable feature representations.

[0107] S3. Determine low-frequency background features, high-frequency detail features, and noise residual features, and stitch together the output feature map. Based on the channel coordinate attention block algorithm, generate an attention weight matrix for dynamic weighting and adjustment. Use global max pooling to extract global dynamic weights and generate an enhanced feature map. Perform dynamic weight allocation based on the enhanced feature map and the weighted feature map, and fuse low-frequency background features and high-frequency detail features to obtain a dynamic fused feature map.

[0108] Preferably, the output feature maps are concatenated, global max pooling is used to extract global dynamic weights, an enhanced feature map is generated, and low-frequency background features and high-frequency detail features are fused to obtain a dynamically fused feature map, including...

[0109] Features are defined based on the denoised image and the corrected brightness image, including low-frequency background features, high-frequency detail features, and noise residual features. These features are then concatenated and combined to output a feature map, represented as follows:

[0110] F low (x,y)=D br (x,y);

[0111] F high (x,y)=D ce (x,y)-F low (x,y);

[0112] F noise (x,y)=D ce (x,y)-Deconv(D ce (x,y));

[0113] F(x,y,c)=concat(F low (x,y),F high (x,y), F noise (x,y))

[0114] Where F low (x,y) represents low-frequency background features, F high (x,y) represents high-frequency detail features, F noise (x,y) represents the noise residual feature, Deconv(D ce (x,y)) represents the transposed convolution of the denoised image, and F(x,y,c) represents the concatenation of feature maps. The concatenation operation concat merges the feature maps in the channel direction, with a total number of channels C = C0. low +C high +C noise C low C high C noise These represent the number of channels for low-frequency background features, high-frequency detail features, and noise residual features, respectively, while C represents the number of channels in the stitched feature map.

[0115] Based on the Channel Coordinate Attention Block (CCAB) algorithm, an attention weight matrix is ​​generated by calculating the statistical importance of feature maps in the channel dimension, as follows:

[0116] W cca (c)=ε(FC2(ReLU(FC1(G avg ))));

[0117]

[0118] Among them W cca (c) represents the dynamic weighting matrix for each channel c of the feature, ε(·) represents the Sigmoid function, FC2(·) represents the second fully connected layer used to restore the channel number dimension, FC1(·) represents the output dimension of the first fully connected layer as C / r (the number of channels is reduced by a factor of r, where r is determined based on historical experience), G acg(c) indicates that H, W, and C represent the height, width, and number of channels of the input feature map, respectively, and F(x,y,c) represents the value of the feature map in the spatial pixel (x,y) and channel c;

[0119] The input feature map is dynamically weighted using an attention weight matrix, and the weighted feature map is output as follows:

[0120] F cca (x,y,c)=F(x,y,c)·W cca (c);

[0121] Where F cca (x,y,c) represents the weighted feature map;

[0122] Based on the feature map, Global Max Pooling (GMP) is used to extract the global maximum value of each channel, generating an initial global feature matrix. This matrix is ​​then transformed through two globally fully connected layers to generate a global dynamic weight matrix, represented as follows:

[0123] W gl (c)=ε(FC2(ReLU(FC1(G max )))));

[0124] Among them W gl (c) represents the global dynamic weight matrix, FC'2 represents the global second fully connected layer used to restore the channel number dimension, FC'1 represents the output dimension of the global first fully connected layer as (C / z, C) (the number of channels is reduced by a factor of z, and r is determined based on historical experience), G max This represents the global max-pooling feature;

[0125] The enhanced feature map is generated using a global dynamic weight matrix, represented as follows:

[0126] F glo (x,y,c)=F(x,y,c)·W gl (c);

[0127] Where F glo (x,y,c) represents the enhanced feature map;

[0128] Dynamic weight allocation is performed based on the enhanced feature map and the weighted feature map, and low-frequency background features and high-frequency detail features are fused to obtain a dynamically fused feature map, represented as follows:

[0129] W c (x,y,c)=ε(F cca (x,y,c)+F glo (x,y,c));

[0130] F en (x,y)=W c(x,y,c)·f high (x,y)(1-W c (x,y,c))·F low (x,y);

[0131] Among them W c (x,y,c) represents the dynamic weight allocation matrix, F en (x,y) represents the dynamically fused feature map.

[0132] By extracting low-frequency information from the denoised image, low-frequency background features are formed to represent the smoothness and large-scale distribution of global illumination. Through transposed convolution, edge information, local details, and target structures of the denoised image are captured, and high-frequency features are enhanced. This solves the problem of over-smoothing that may occur during the denoising process, and preserves clear boundary and texture information. The noise residual features, by specially separating the special components of residual noise, intuitively guide the subsequent dynamic weighting and fusion modules to further reduce noise, while ensuring that the edges are not overly smoothed or the residual spreads during the denoising process. After the three features are stitched together, the generated stitched feature map integrates the information of different features in the channel dimension, realizing a comprehensive representation of the image in the spatial and frequency domains, and improving the optimization capability of the subsequent attention mechanism.

[0133] The CCAB algorithm is used to calculate the statistical importance of feature maps and generate a dynamic weighted matrix, which adjusts the feature weights of each channel. Channel weighting effectively suppresses the redundancy that may exist in specific channels in the noise residual, thereby reducing the interference of useless information on subsequent processing and improving the sensitivity to meaningful signals. The importance of channels is extracted by global average pooling in the channel dimension and generated after transformation through two fully connected layers. The weight value of each channel is dynamically assigned. During the feature map weighting process, the dynamic adjustment of channel weights makes the response of each channel more consistent with different partitions of the image, improving the flexibility of the model in switching between local and global modes. For example, it performs more stably when processing bright backgrounds and edge regions with more details.

[0134] Global max pooling (GMP) is used to extract the global maximum value of each channel, which is then mapped to a global dynamic weight matrix through two fully connected layers to further optimize channel feature allocation. GMP is used to capture the most salient features of each channel in the entire image, thereby helping to emphasize the most important information regions in the infrared image. Feature enhancement is particularly important for low-frequency and brightness correction regions, because the background often needs to be adjusted at a global scale to improve illumination smoothness. Unlike CCAB, which focuses more on the importance of local channels, GMP emphasizes global saliency, ensuring that the generator does not ignore the correlation of the overall region when optimizing brightness. The addition of global dynamic weights builds a highly consistent global foundation for subsequent dynamic fusion.

[0135] By dynamically weighting and fusing low-frequency and high-frequency features, the fusion result can achieve a balanced feature distribution in different scenarios (such as high-contrast scenarios and low-contrast uniform scenarios). Low-frequency background features, noise residual features, and high-frequency detail features may inherently contain interference (such as some noise residuals in high frequencies). The dynamic weighting mechanism reduces the impact of this potential problem by suppressing non-complementary characteristics, thereby optimizing the quality of the final feature map.

[0136] S4. Multi-scale sampling operation is used to generate feature maps of different scales based on the dynamic fusion feature map. The dynamic fusion feature map is then re-input into the generator to generate an updated input image. The encoder and decoder are then used to fuse the acquired optimized denoised image and corrected brightness image to obtain an optimized enhanced image.

[0137] Preferably, multi-scale sampling operations are used to generate feature maps of different scales based on the dynamically fused feature maps. These dynamically fused feature maps are then re-inputted into the generator to generate an updated input image, including...

[0138] Based on the dynamically fused feature map, multi-scale sampling operations are used to generate feature maps of different scales;

[0139] Based on the determined fusion weights, the final multi-scale features are obtained by weighted fusion of feature maps generated from multiple scales, as follows:

[0140]

[0141] Where F fu (x,y) represents the multi-scale feature map, S represents the total number of scales, and β s Indicates the fusion weight. This represents the dynamic fusion feature map at scale s.

[0142] By dynamically fusing feature maps and generating feature maps of different resolutions through multi-scale sampling, this multi-scale feature extraction strategy improves the ability to capture image information. It enables multi-scale feature maps to optimize the balance between global scene representation and local detail enhancement in sample adaptation. Through a weight allocation mechanism, the contribution ratio of each scale feature is optimized according to its importance in a specific scene, eliminating potential conflicts and making the final fusion result more coordinated and accurate.

[0143] Furthermore, the encoder and decoder are reused to perform data fusion on the acquired optimized denoised image and brightness-corrected image to obtain an optimized and enhanced image, including...

[0144] The dynamically fused feature map is re-inputted into the generator, and then concatenated with the high-frequency and low-frequency feature images to generate an updated input image, represented as follows.

[0145] I'gen (x,y,c)=concat(I hf (x,y),I cf (x,y));

[0146] Among them I' gen (x,y,c) represents the updated input image of the generator, I hf (x,y) represents the high-frequency feature image, I cf (x,y) represents the low-frequency feature image;

[0147] Based on the updated input image, the encoder and decoder are reused to obtain the decoded outputs of different convolutional layers. A denoising filter is then used to obtain an optimized denoised image and a corrected brightness image. Furthermore, data fusion is performed using the sigmoid function to obtain an optimized and enhanced image, represented as:

[0148] I la (x,y)=D′ ca (x,y)+sigmoid(D′ br (x,y))

[0149] Where I la (x,y) represents the optimized and enhanced image, D' ca (x,y) represents the optimized denoised image, D' br (x,y) represents the optimized and corrected brightness image.

[0150] Dynamically fused feature maps, being the result of multi-scale features and weight allocation, possess strong integrative properties, reflecting a comprehensive expression of local image details and global brightness. Reintroducing high-frequency and low-frequency feature images further supplements the independently extracted detail information, enhancing the characteristic data of local edge details (high-frequency features) and wide-area illumination balance (low-frequency features), thereby expanding the dimensionality of the input data. Since dynamically fused feature maps may lose some high-frequency or low-frequency details during multiple weight allocation and reconstruction processes, high-frequency, low-frequency, and dynamically fused features are separated into representations of different scales and directions in the encoder's multi-layer convolutional operations. Higher-level convolutions extract global features (such as background illumination), while lower-level convolutions focus on local features (such as edge details), thus achieving a more comprehensive representation. Feature extraction and dynamic fusion of features are used as part of the encoder input. The optimization results of multi-scale weight allocation are integrated through convolution transformation, providing the encoder with a more reliable global background reference. The decoder's ability to suppress residual noise in high-frequency features is significantly improved. Through the denoising result after decoding, noise is effectively removed while edge and texture characteristics are preserved. The sigmoid nonlinear function prevents the denoising result from directly covering the brightness correction result by smoothing the output range, ensuring that the allocation of denoising and brightness correction in the enhanced image is more reasonable. The optimized enhanced image has both clear target details (high-frequency part is preserved) and uniform and smooth background brightness (low-frequency part is optimized), ultimately achieving a good balance between target and scene integrity.

[0151] S5, iterate and re-output the optimized and enhanced image, and analyze it by comparing the peak signal-to-noise ratio and structural similarity index with the corresponding threshold;

[0152] Preferably, the optimized and enhanced image is iteratively re-output, and the peak signal-to-noise ratio and structural similarity indices are compared and analyzed against corresponding thresholds, including:

[0153] Based on the iterative re-output of the dynamically fused feature map, the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) of the optimized and enhanced image are calculated.

[0154] Based on the reference real images in the validation set data, the signal-to-noise ratio threshold and similarity threshold are determined. If the peak signal-to-noise ratio of the optimized and enhanced image is greater than the signal-to-noise ratio threshold and the structural similarity index of the optimized and enhanced image is greater than the similarity threshold, the iteration is stopped and the optimized and enhanced image of the current iteration is used as the final processed image.

[0155] By guiding the model through PSNR to optimize the output image at a global scale, especially for large target areas or large background scenes, its highly robust noise evaluation mechanism can effectively avoid pseudo-optimization caused by the model over-amplifying local noise details. The final optimized image determined by SSIM retains the local structural information maintained in multiple iterations of the dynamic feature map, preventing excessive smoothing of pixel values ​​from negatively impacting infrared image applications (such as edge target recognition). The joint determination of the two thresholds not only promotes the balance of collaborative optimization between the two tasks, but also avoids either task being over-prioritized and damaging the final image quality. Through multiple quality monitoring of global and local feature matching, the enhanced image terminated by PSNR and SSIM determination is more in line with the needs of actual applications.

[0156] This embodiment also provides an artificial intelligence-based infrared image denoising system, including,

[0157] The data preprocessing module acquires infrared image data and extracts information from different frequencies of the image to determine high-frequency feature maps and low-frequency background feature maps.

[0158] The generator module concatenates the initial multi-channel input image, extracts multi-scale semantic features, and gradually restores the image resolution using a two-dimensional transposed convolution operation. It also uses a denoising filter to further process and reduce noise, and corrects the brightness component.

[0159] The loss calculation and optimization module defines a comprehensive loss function and iteratively optimizes it through gradient backpropagation and Adam optimizer to finally output a stable denoised image and a brightness-corrected image.

[0160] The dynamic feature extraction module extracts low-frequency background features, high-frequency detail features, and noise residual features, splices them into a multi-channel feature map, generates an enhanced feature map through global max pooling, and generates a dynamic fused feature map.

[0161] The multi-scale sampling enhancement module generates feature maps of different scale ranges by dynamically fusing feature maps, and optimizes the denoised image and brightness correction image through encoder and decoder to generate an optimized and enhanced image.

[0162] The image quality assessment module calculates the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) of the optimized and enhanced images, and performs comparative analysis with the real images in the validation set.

[0163] This embodiment also provides a computer device applicable to the case of an artificial intelligence-based infrared image denoising method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the artificial intelligence-based infrared image denoising method proposed in the above embodiment.

[0164] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0165] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the artificial intelligence-based infrared image denoising method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0166] In summary, this invention effectively preserves spatial structural information such as edges and textures by transforming infrared images into high-frequency and low-frequency feature decomposition and extracting high-frequency features based on Gaussian high-pass filters. The attention mechanism weights the importance of these features and enhances noise detection. For the image background, the attention module identifies areas with excessive brightness fluctuations. Residual characteristics synchronize the transmission of secondary features during noise removal, preventing feature mutations or information loss during decoding. The joint enhancement of the denoising filter and brightness correction nonlinearly adjusts the high-frequency noise intensity, suppressing residual noise while preserving high-frequency sharpening characteristics. Gaussian parameters ensure controllable noise processing intensity, avoiding excessive erasure of key details. Joint optimization of the comprehensive loss function improves task synergy, and supervised regularization strengthens the physical properties of the image from the perspective of gradient constraints. By extracting low-frequency information from the denoised image, low-frequency background features are formed to represent the smooth features and large-scale distribution of global illumination. Dynamic weight allocation fuses low-frequency and high-frequency features, ensuring balanced feature distribution across different scenarios.

[0167] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An infrared image denoising method based on artificial intelligence, characterized in that, The method comprises the following steps: Collecting infrared image data to extract frequency domain spatial features, performing Gaussian filtering and restoring to image space to determine high-frequency feature maps, extracting image background information to determine low-frequency feature maps, and splicing the high-frequency feature maps with the generator, using an encoder based on a convolutional transpose self-attention mechanism to perform two-dimensional transpose convolution operation, using residual connection to obtain a joint output of the denoised high-frequency feature map, cooperating with a denoising filter and correcting the brightness component to obtain a denoised and brightness-corrected image; Based on the supervised regularization mechanism, the loss of the denoised and brightness-corrected image is calculated, and the comprehensive loss is determined according to the physical consistency loss, the loss of the denoising task and the loss of the brightness correction task, the generator parameters are iteratively optimized through gradient back propagation, and a stable denoised and brightness-corrected image is output. Determine the low-frequency background feature, the high-frequency detail feature and the noise residual feature, splice the output feature map, generate an attention weight matrix according to the channel coordinate attention block algorithm for dynamic weighting adjustment, use global maximum pooling to extract global dynamic weights, generate an enhanced feature map, and perform dynamic weight distribution according to the enhanced feature map and the weighted feature map to obtain a dynamic fusion feature map. Using multi-scale sampling operation, different scale feature maps are generated according to the dynamic fusion feature map, the generator is re-inputted according to the dynamic fusion feature map, an updated input image is generated, and the obtained optimized denoised image and corrected brightness image are data fused by using the encoder and the decoder to obtain an optimized enhanced image. Iteratively output the optimized enhanced image, and compare and analyze the peak signal-to-noise ratio and the structural similarity index with the corresponding threshold value. 2.The AI-based infrared image denoising method of claim 1, wherein: The method comprises the following steps: Collecting infrared image data to extract frequency domain spatial features, splicing the high-frequency feature maps with the generator, using an encoder based on a convolutional transpose self-attention mechanism, cooperating with a denoising filter and correcting the brightness component, Collecting infrared image data, converting to frequency domain space using two-dimensional discrete Fourier transform, calculating the amplitude based on the amplitude and phase information of the frequency domain space data at the corresponding frequency; Respectively using a Gaussian high-pass filter as a filtering reference to determine a high-frequency filtered frequency domain image, and using two-dimensional inverse Fourier transform to restore to image space to obtain a high-frequency feature map; Using a Gaussian low-pass filter to extract image background information, applying a low-frequency filter to the image in the frequency domain, and using 2D-IDFT to restore to image space to obtain a low-frequency feature map; The extracted high-frequency image and low-frequency image are verified with the original image respectively, and a multi-channel input image is obtained by splicing; Based on the encoder based on the convolutional transpose self-attention mechanism CTASB, wherein the encoder extracts features from the multi-channel input image through a convolutional layer and performs feature weighting to obtain attention features, and the decoder performs two-dimensional transpose convolution operation on the attention features to obtain decoding outputs of different convolutional layers. The residual connection method is used to feed back part of the features of the encoder to the decoding stage directly to obtain a joint output of a denoised high-frequency feature map, a denoising filter is used for denoising processing based on the denoised high-frequency feature map, and a normalization activation function is used to correct the luminance component to obtain an image after denoising and luminance correction. 3.The AI-based infrared image denoising method of claim 2, wherein: Loss calculation is performed on the image after denoising and luminance correction, generator parameters are iteratively optimized through gradient back propagation, and a stable denoised and luminance corrected image is output, including, Loss calculation is performed based on a supervised regularization mechanism, regularization term calculation is performed for the first-order gradient of the denoised image and the second-order gradient of the corrected luminance image, respectively, wherein the first-order gradient of the denoised image is subjected to L1 norm constraint, the gradient amplitude sum is calculated, and the second-order gradient of the corrected luminance image is subjected to L2 norm constraint; Based on the input image data, the denoised image and the corrected luminance image are combined to calculate the physical consistency loss, and the loss of the denoising task, the loss of the luminance correction task, the supervised regularization term and the physical consistency loss are jointly optimized to determine the comprehensive loss; The gradient is calculated through gradient back propagation according to the comprehensive loss, the gradient is optimized through the Adam optimizer, the generator parameters are updated, and when the comprehensive loss no longer decreases significantly in the continuous iteration process, the iteration is stopped and the generator parameters are output, and a stable denoised image and a corrected luminance image are output. 4.The AI-based infrared image denoising method of claim 3, wherein: The output feature map is spliced, global dynamic weights are extracted using global maximum pooling, enhanced feature maps are generated, and dynamic fusion feature maps are obtained by fusing low-frequency background features and high-frequency detail features, including, Based on the denoised image and the corrected luminance image, features including low-frequency background features, high-frequency detail features and noise residual features are defined and spliced to output feature maps; According to the channel coordinate attention block CCAB algorithm, an attention weight matrix is generated by calculating the statistical importance of the feature map in the channel dimension, the input feature map is dynamically weighted and adjusted using the attention weight matrix, and a weighted feature map is output. According to the feature map, a global maximum value of each channel is extracted using global maximum pooling GMP to generate an initial global feature matrix, which is transformed through two global fully connected layers to generate a global dynamic weight matrix, and the global dynamic weight matrix is used to generate enhanced feature maps. According to the enhanced feature map and the weighted feature map, dynamic weight distribution is performed to fuse low-frequency background features and high-frequency detail features to obtain dynamic fusion feature maps. 5.The AI-based infrared image denoising method of claim 4, wherein: The multi-scale sampling operation is used to generate feature maps of different scales based on the dynamic fusion feature map, and the generator is re-input based on the dynamic fusion feature map to generate an updated input image, including, According to the dynamic fusion feature map, a multi-scale sampling operation is used to generate feature maps of different scales, a fusion weight is determined, and a final multi-scale feature is obtained by weighted fusion of the feature maps generated by multiple scales. 6.The AI-based infrared image denoising method of claim 5, wherein: The encoder and the decoder are used again to fuse the obtained optimized denoised image and the corrected luminance image to obtain an optimized enhanced image, including, According to the dynamic fusion feature map re-input generator, the high-frequency feature image and the low-frequency feature image are spliced to generate an updated input image; According to the updated input image, the decoder and the encoder are used again to obtain the decoding output of different convolution layers, and the denoising filter is used to obtain the optimized denoising image and the corrected brightness image, and further data fusion is performed through the sigmoid function to obtain the optimized enhancement image. 7.The AI-based infrared image denoising method of claim 6, wherein: The iterative re-output optimization enhancement image is output, and according to the comparison and analysis of the peak signal-to-noise ratio and the structural similarity index with the corresponding threshold value, According to the iterative re-output of the dynamic fusion feature map, the peak signal-to-noise ratio PSNR and the structural similarity index SSIM of the optimized enhancement image are calculated, and according to the reference true image of the verification set data, the signal-to-noise ratio threshold and the similarity threshold are determined. If the peak signal-to-noise ratio of the optimized enhancement image is greater than the signal-to-noise ratio threshold, and the structural similarity index of the optimized enhancement image is greater than the similarity threshold, the iteration is stopped, and the optimized enhancement image of the current iteration is taken as the final processing image.

8. The infrared image denoising system based on artificial intelligence according to any one of claims 1 to 7, characterized in that: Including, The data preprocessing module collects infrared image data and extracts different frequency information of the image to determine the high-frequency feature map and the low-frequency background feature map; The generator module splices to generate an initial multi-channel input image, extracts multi-scale semantic features, and gradually restores the image resolution by using two-dimensional transpose convolution operation, further processes noise reduction by using a denoising filter, and corrects the brightness component; The loss calculation optimization module defines a comprehensive loss function, and iteratively optimizes the final output stable denoising image and brightness correction image through gradient back propagation and Adam optimizer; The dynamic feature extraction module extracts low-frequency background features, high-frequency detail features and noise residual features, splices them into a multi-channel feature map, generates an enhanced feature map through global maximum pooling, and generates a dynamic fusion feature map; The multi-scale sampling enhancement module generates feature maps of different scale ranges through the dynamic fusion feature map, optimizes the denoising image and the brightness correction image through the encoder and the decoder, and generates the optimized enhancement image; The image quality evaluation module calculates the peak signal-to-noise ratio PSNR and the structural similarity index SSIM of the optimized enhancement image, and compares and analyzes it with the reference true image of the verification set. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to realize the steps of the artificial intelligence-based infrared image denoising method of any one of claims 1-7.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the artificial intelligence-based infrared image denoising method of any one of claims 1-7.

Images