Infrared non-uniformity correction method, correction method based on lightweight U-MISO network

Abstract

This invention relates to an image correction method. Addressing the technical problems of existing infrared non-uniformity correction methods, such as the tendency for reference-source-based methods to produce ghosting in striped images and the inability of deep learning-based methods to achieve real-time processing speeds suitable for engineering applications, this invention provides an infrared non-uniformity correction method based on a lightweight U-MISO network. This method first downsamples the non-uniform noise image twice, then uses multiple convolutions, residual networks, deconvolutions, multi-scale fusion modules, feature attention modules, and overlay processes to extract features from the non-uniform noise image multiple times. Finally, it accurately obtains the noise features of the global image. Subtracting these noise features from the non-uniform noise image yields the corrected image. This method significantly improves algorithm speed while effectively removing non-uniform noise.

Description

Technical Field

[0001] This invention relates to an image correction method, specifically an infrared non-uniformity correction method and a correction method based on a lightweight U-MISO network. Background Technology

[0002] Currently, infrared non-uniformity correction (NUC) methods are mainly divided into two categories. One is based on a reference radiation source, using a uniformly temperatured blackbody to calculate fixed correction coefficients, such as one-point correction, two-point correction, and multi-point correction methods. The prerequisite for using the one-point correction method is that the ambient temperature fluctuation of the entire infrared imaging system is small, and the infrared radiation energy incident on the system by the external target is close to the ambient temperature. While meeting these environmental conditions, the gain coefficients of each pixel of the detector also differ little. Generally, the one-point correction method only corrects the response bias. When the infrared imaging system is under the proposed assumptions, its response non-uniformity is mainly caused by fixed additive background noise. However, in actual measurements, the infrared imaging system not only has additive background noise but also multiplicative and additive non-uniform noise. The causes of this noise can be attributed to the difference between the DC bias and gain of the infrared focal plane detector. Therefore, in practical applications, simply using the one-point correction method to correct the non-uniformity of the infrared imaging system usually results in significant errors. Based on this, to correct the shortcomings of the one-point correction method, a two-point correction method was developed, which simultaneously corrects both gain and bias. However, under real-world operating conditions, considering the nonlinearity of the detector response, the two-point correction method cannot meet the requirements of a wider temperature range, thus leading to the development of a multi-point correction method. Essentially, the difference between the multi-point and two-point correction methods is not significant. In practice, the multi-point correction method involves selecting multiple different temperature points and performing two-point corrections between adjacent temperature points. By repeatedly performing two-point corrections between different points, a reasonable correction over a wider temperature range is achieved, and multi-segment linear simulations are performed on this nonlinear output.

[0003] However, due to the inherent time drift characteristics of the detector, reference source-based methods require periodic interruptions for recalibration. To overcome this limitation, researchers have proposed many scene-based NUC methods with different operating principles. These methods can be further divided into multi-frame and single-frame methods. These methods heavily rely on inter-frame scene motion or parameter selection, and are prone to ghosting in the striped images.

[0004] In recent years, with the continuous increase in image data volume and the continuous improvement of hardware performance, deep learning has been successfully applied to image processing problems such as computer vision. Driven by the development of this technology, researchers have proposed some methods for removing stripe non-uniform noise, such as CNN-based methods: SNRCNN, SNRWDNN, ICSR, WDNN, DLS-NUC, and ICSRN. Although these methods are superior to classical stripe removal methods in both quantitative and qualitative evaluation, they also have their limitations. SNRCNN and ICSR cannot completely remove strong stripe non-uniform noise. SNRWDNN removes vertical edges in the image as stripe noise, resulting in the loss of scene details. Although DLS-NUC solves the problem of incomplete removal of non-uniform noise, it reduces contrast in some infrared image processing. WDNN improves the algorithm's adaptability to noise intensity and distribution; however, like SNRCNN and ICSR, this algorithm has a low peak signal-to-noise ratio when encountering high-intensity noise. Most importantly, the single-frame speed of the above deep learning-based methods for processing infrared images cannot achieve engineering-scale real-time performance, making them impractical for real-world applications. Summary of the Invention

[0005] This invention addresses the technical problems of existing infrared non-uniformity correction methods, such as the tendency of reference radiation source-based methods to produce ghosting in striped images and the inability of deep learning-based methods to achieve real-time processing speeds for single frames of infrared images in engineering applications. It provides an infrared non-uniformity correction method based on a lightweight U-MISO network.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] An infrared non-uniformity correction method, characterized by the following steps:

[0008] S1, perform downsampling twice on the non-uniform noise image I1 to obtain images I2 and I3;

[0009] S2, features F are obtained by performing convolution on the non-uniform noise image I1. 01 ; respectively through the shallow feature extraction module f SFEM Shallow feature extraction is performed on images I2 and I3 to obtain shallow feature F. 02_SFEM and shallow features F 03_SFEM ;

[0010] S3, using a residual network to process feature F 01 The feature F is obtained through processing. 01_RB ;

[0011] S4, using convolution on feature F 01_RBPreliminary feature extraction is performed to obtain feature F. 01_AB Then, the feature attention module is used to process feature F. 01_AB and feature F 02_SFEM Feature extraction is performed to obtain feature F. 02_FA Then, the feature F is processed through a residual network. 02_FA Feature extraction is performed to obtain feature F. 02_RB ;

[0012] S5, using convolution on feature F 02_RB Preliminary feature extraction is performed to obtain feature F. 02_AB Then, the feature attention module is used to process feature F. 02_AB and feature F 03_SFEM Feature extraction is performed to obtain feature F. 03_FA Then, the feature F is processed through a residual network. 03_FA Feature extraction is performed to obtain feature F. 03_RB ;

[0013] S6, using multi-scale fusion modules of two different scales to respectively process feature F 01_RB Feature F 02_RB Feature F 03_RB The feature F is obtained through processing. 01_MF and feature F 02_MF The scales of the two multi-scale fusion modules are equal to the scales of the non-uniform noise image I1 and image I2, respectively.

[0014] S7, for feature F 03_RB The feature F is obtained by sequentially processing the data through a residual network and deconvolution. 03_CT ;

[0015] S8, for feature F 03_CT With feature F 02_MF By stacking along the feature channel dimension, we obtain feature F. cat2_3_1 ;

[0016] S9, features F are sequentially processed through convolution, residual network, and deconvolution. cat2_3_1 The feature F is obtained through processing. 02_CT ;

[0017] S10, for feature F 01_MF With feature F 02_CT 1. The noise features F of the global image are obtained by stacking them along the feature channel dimension and then sequentially passing them through convolution, residual network, and convolution. noise ;

[0018] S11, Subtract noise features F from the non-uniform noise image I1 noise The corrected image output is obtained.

[0019] Furthermore, in step S1, the downsampling specifically involves downsampling by a factor of 2 using the nearest neighbor method;

[0020] In steps S2 and S9, the convolution is a 3×3 convolution with a stride of 1;

[0021] In step S2, the shallow feature extraction module f SFEM It includes two 3×3 convolutions, three 1×1 convolutions, and a channel concatenation structure. The three 1×1 convolutions are connected in sequence. The output of one 3×3 convolution is connected to the input of the first 1×1 convolution. Another 3×3 convolution is connected between the first and second 1×1 convolutions. The channel concatenation structure is connected between the second and third 1×1 convolutions. The other input of the channel concatenation structure is used to input image I2 or image I3.

[0022] Furthermore, in steps S3, S4, S5, S7, S9, and S10, the residual network is composed of 8 residual blocks connected in series; in steps S4 and S5, the convolutions are all 2×2 convolutions with a stride of 2; in steps S9 and S10, the convolutions are all 3×3 convolutions with a stride of 1.

[0023] Further, in step S4, the feature attention module includes a matrix dot product structure, a 3×3 convolution, and a matrix summation structure. The 3×3 convolution is connected between the output of the matrix dot product structure and one input of the matrix summation structure. The two inputs of the matrix dot product structure are respectively connected to the output of the convolution and the shallow feature extraction module f corresponding to the non-uniform noise image I2. SFEM The matrix points and the other input end of the structure are connected to the output end of the convolution;

[0024] In step S5, the feature attention module includes a matrix dot product structure, a 3×3 convolution, and a matrix summation structure. The 3×3 convolution is connected between the output of the matrix dot product structure and one input of the matrix summation structure. The two inputs of the matrix dot product structure are respectively connected to the output of the convolution and the shallow feature extraction module f corresponding to the non-uniform noise image I3. SFEM The matrix points and the other input end of the structure are connected to the output end of the convolution.

[0025] Further, in step S6, the multi-scale fusion module includes three resize modules, one 1×1 convolution, one 3×3 convolution, and one channel concatenation structure. The three inputs of the channel concatenation structure are respectively connected to the three resize modules, and the outputs of the channel concatenation structure are sequentially connected to the 1×1 convolution and the 3×3 convolution. The output of the 3×3 convolution serves as the multi-scale fusion module f. MFThe output of the module and the inputs of the three resize modules serve as the three inputs of the multi-scale fusion module, respectively receiving features F. 01_RB Feature F 02_RB and feature F 03_RB .

[0026] Furthermore, in steps S7 and S9, the deconvolution specifically involves performing a double upsampling process through deconvolution.

[0027] This invention also provides a calibration method based on a lightweight U-MISO network, characterized by the following steps:

[0028] S1, Construct a lightweight U-MISO network, wherein the lightweight U-MISO network is equipped with the infrared non-uniformity correction method as described in any one of claims 1 to 6;

[0029] S2, the lightweight U-MISO network is trained using multiple sets of known non-uniform noise images and corresponding clean infrared images as training sets, resulting in the trained lightweight U-MISO network.

[0030] S3. Input any non-uniform noise image into the trained lightweight U-MISO network, and use the output of the trained lightweight U-MISO network as the correction result.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. This invention proposes an infrared non-uniformity correction method. First, the non-uniform noise image is downsampled twice. Then, through multiple convolutions, residual networks, deconvolutions, multi-scale fusion modules, feature attention modules, and overlay processes, features are extracted multiple times from the non-uniform noise image. Finally, the noise features of the global image are accurately obtained. Subtracting these noise features from the non-uniform noise image yields the corrected image. This method significantly improves algorithm speed while effectively removing non-uniform noise. Furthermore, this invention can be applied to any infrared imaging-related technical field, such as infrared security monitoring and infrared human body temperature measurement.

[0033] 2. The present invention also provides a correction method based on a lightweight U-MISO network. A lightweight U-MISO network incorporating the above correction method is constructed as a deep learning network model. By combining the above correction method and deep learning technology, rapid correction of infrared non-uniformity can be achieved, effectively improving the efficiency and accuracy of correction. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of an embodiment of the infrared non-uniformity correction method of the present invention;

[0035] Figure 2 The shallow feature extraction module f in the embodiment of the infrared non-uniformity correction method of the present invention SFEM A schematic diagram;

[0036] Figure 3 The feature attention module f in the embodiment of the infrared non-uniformity correction method of the present invention FA A schematic diagram;

[0037] Figure 4 This is a schematic diagram of the multi-scale fusion module in an embodiment of the infrared non-uniformity correction method of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0039] This invention proposes an infrared non-uniformity correction method based on a lightweight U-MISO network and establishes a U-MISO network model. The specific algorithm in the U-MISO network model is as follows, and the following algorithm can also be used as an infrared non-uniformity correction method on its own.

[0040] Using the original non-uniform noise image I1 with stripes as the input of the U-MISO network model, infrared non-uniformity correction is performed on the non-uniform noise image I1 according to the following method: Step 1, the non-uniform noise image I1 is downsampled by 2 times using the nearest neighbor method to obtain the non-uniform noise image I2. Similarly, I2 is downsampled by 2 times using the nearest neighbor method to obtain the uniform noise image I3.

[0041] Step 2: Extract features from the non-uniform noise images I1, I2, and I3 from Step 1.

[0042] Step 2.1: Use a 3×3 convolution f1 with a stride of 1 to extract features from the non-uniform noise image I1, obtaining feature F. 01 :

[0043] F 01 =f1(I1)

[0044] Simultaneously, the non-uniform noise image I2 and the non-uniform noise image I3 are processed by the shallow feature extraction module (SFEM) respectively. SFEM Perform shallow feature extraction to obtain the corresponding shallow feature F. 02_SFEMand shallow features F 03_SFEM :

[0045] F 02_SFEM =f SFEM (I2)

[0046] F 03_SFEM =f SFEM (I3)

[0047] Among them, such as Figure 2 As shown, the shallow feature extraction module f SFEM It includes two 3×3 convolutions, three 1×1 convolutions, and a channel stitching structure with a skip connection. The three 1×1 convolutions are connected in sequence, one 3×3 convolution is connected before the first 1×1 convolution, and another 3×3 convolution is connected between the first and second 1×1 convolutions. The channel stitching structure is connected between the second and third 1×1 convolutions. The channel stitching structure is used to stitch together the output of the second 1×1 convolution and the non-uniform noise image I2 or the non-uniform noise image I3.

[0048] Step 2.2, for feature F 01 Furthermore, a residual network f consisting of 8 cascaded residual blocks is used. RB Processing yields feature F 01_RB :

[0049] F 01_RB =f RB (F 01 )

[0050] A 3×3 convolution f2 with a stride of 2 is used to apply the feature F 01_RB Preliminary feature extraction is performed to obtain feature F. 01_AB :

[0051] F 01_AB =f2(F 01_RB )

[0052] For feature F 01_AB and feature F 02_SFEM Through Feature Attention Module (FA) f FA Feature extraction is performed to obtain feature F. 02_FA :

[0053] F 02_FA =f FA (F 01_AB ,F 02_SFEM )

[0054] Among them, such as Figure 3 As shown, the feature attention module f FAIt includes a matrix dot product structure, a 3×3 convolution, and a matrix summation structure connected in sequence. The 3×3 convolution is connected between the output of the matrix dot product structure and one input of the matrix summation structure. The two inputs of the matrix dot product structure are respectively connected to the output of convolution f2 and the shallow feature extraction module f corresponding to the non-uniform noise image I2. SFEM The other input of the matrix points and structure is connected to the output of convolution f2.

[0055] Then, through a residual network f consisting of 8 residual blocks connected in series... RB For feature F 02_FA Feature extraction is performed to obtain feature F. 02_RB :

[0056] F 02_RB =f RB (F 02_FA )

[0057] Step 2.3, synchronized with step S2.2:

[0058] A 3×3 convolution f3 with a stride of 2 is used to apply the feature F 02_RB Preliminary feature extraction is performed to obtain feature F. 02_AB :

[0059] F 02_AB =f2(F 02_RB )

[0060] For feature F 02_AB and feature F 03_SFEM Through Feature Attention Module (FA) f FA Feature extraction is performed to obtain feature F. 03_FA :

[0061] F 03_FA =f FA (F 02_AB ,F 03_SFEM )

[0062] Among them, the feature attention module f FA It includes a matrix dot product structure, a 3×3 convolution, and a matrix summation structure connected in sequence. The 3×3 convolution is connected between the output of the matrix dot product structure and one input of the matrix summation structure. The two inputs of the matrix dot product structure are respectively connected to the output of convolution f3 and the shallow feature extraction module f corresponding to the non-uniform noise image I3. SFEM The other input of the matrix points and structure is connected to the output of convolution f3.

[0063] Then, through a residual network f consisting of 8 residual blocks connected in series... RB For feature F 03_FA Feature extraction is performed to obtain feature F.03_RB :

[0064] F 03_RB =f RB (F 03_FA )

[0065] Step 2.4, respectively, for feature F 01_RB Feature F 02_RB Feature F 03_RB Through the multi-scale fusion module (MF)f MF1 and multi-scale fusion module f MF2 The feature F is obtained through processing. 01_MF and feature F 02_MF :

[0066] F 01_MF =f MF1 (F 01_RB ,F 02_RB ,F 03_RB )

[0067] F 02_MF =f MF2 (F 01_RB ,F 02_RB ,F 03_RB )

[0068] Among them, such as Figure 4 As shown, the multi-scale fusion module (MF) f MF1 and multi-scale fusion module f MF2 Each module includes three resize modules: a 1×1 convolution, a 3×3 convolution, and a channel concatenation structure, as well as a multi-scale fusion module f. MF The channel stitching structure has three inputs connected to three resize modules, and its outputs are connected to a 1×1 convolution and a 3×3 convolution. The output of the 3×3 convolution serves as the multi-scale fusion module f. MF The output of the module and the inputs of the three Resize modules serve as the multi-scale fusion module f. MF The three input terminals respectively receive feature F 01_RB Feature F 02_RB and feature F 03_RB Multi-scale fusion module (MF) MF1 and multi-scale fusion module f MF2 The different scales of the Resize module in the model result in different F values. 01_MF and F 02_MF The scale size is equal to that of the non-uniform noise image I1 and the non-uniform noise image I2, respectively.

[0069] Step 2.5, transfer feature F 03_RB The residual network f, consisting of 8 cascaded residual blocks, is used.RB The feature F is obtained through processing. 03_RB2 :

[0070] F 03_RB2 =f RB (F 03_RB )

[0071] Then through deconvolution f CONV_T For F 03_RB2 Perform a 2x upsampling process to obtain feature F 03_CT :

[0072] F 03_CT =f CONV_T (F 03_RB2 )

[0073] Step 2.6, for feature F 03_CT With feature F 02_MF The features are stacked along the feature channel dimension and then obtained by a 3×3 convolution f4 with a stride of 1. cat2_3_1 :

[0074] F cat2_3_1 =f4(concat(F 03_CT ,F 02_MF ))

[0075] Step 2.7, transfer feature F cat2_3_1 The residual network f, consisting of 8 cascaded residual blocks, is used. RB The feature F is obtained through processing. cat2_3_RB :

[0076] F cat2_3_RB =f RB (F cat2_3_1 )

[0077] Then through deconvolution f CONV_T For feature F cat2_3_RB Perform a 2x upsampling process to obtain feature F 02_CT :

[0078] F 02_CT =f CONV_T (F cat2_3_RB )

[0079] Step 2.8, transfer feature F 02_CT With F 01_MF The features are stacked along the feature channel dimension and then obtained by a 3×3 convolution f5 with a stride of 1. cat1_2_1 :

[0080] F cat1_2_1 =f5(concat(F 02_CT ,F 01_MF ))

[0081] Step 2.9, transfer feature F cat1_2_1 First, a residual network f consisting of 8 residual blocks connected in series is used. RB After processing, feature extraction is performed using a 3×3 convolution with a stride of 1, f6, to obtain the noise features F of the global image. noise :

[0082] F noise =f5(F cat1_2_1 )

[0083] Step 2.10: Compare the non-uniform noise image I1 with the noise features F noise Subtracting the two results in the final corrected output image.

[0084] output = I1 - F noise .

[0085] Therefore, the infrared non-uniformity correction method based on the lightweight U-MISO network is based on the above-mentioned infrared non-uniformity correction method, and the specific steps are as follows:

[0086] (1) Construct a lightweight U-MISO network as a deep learning network model. The lightweight U-MISO network is equipped with the infrared non-uniformity correction method described in steps 1 to 2.10 above.

[0087] (2) Multiple sets of known non-uniform noise images containing non-uniform noise data and corresponding clean infrared images are used as the training set for the lightweight U-MISO network.

[0088] When selecting the training set, the non-uniformity noise model published by Z. He, Y. Cao, Y. Dong et al. in "Single-image-based nonuniformity correction of uncooled long-wave infrared detectors" (pages 155-164 of *A Deep-learning Approach*, 2018) can be used to generate a paired dataset of non-uniformly noisy images and clean infrared images. The lightweight U-MISO network is trained, and the parameter set Θ after training is determined, resulting in the trained lightweight U-MISO network. The loss function Loss of the lightweight U-MISO network is as follows:

[0089] Loss = -SSIM

[0090] SSIM stands for S-SIM, which is used to evaluate the similarity between two images.

[0091] (3) Input any non-uniform noise image into the trained lightweight U-MISO network, and apply the trained parameter set Θ to use the output of the trained lightweight U-MISO network as the correction result of the non-uniform noise image.

[0092] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for infrared non-uniformity correction, characterized in that, Includes the following steps: S1, the non-uniform noise image I1 is downsampled twice to obtain images I2 and I3; S2, features F are obtained by performing convolution on the non-uniform noise image I1. 01 ; respectively through the shallow feature extraction module f SFEM Shallow feature extraction is performed on images I2 and I3 to obtain shallow feature F. 02_SFEM and shallow features F 03_SFEM ; S3, using a residual network to process feature F 01 The feature F is obtained through processing. 01_RB ; S4 uses convolution on feature F 01_RB Preliminary feature extraction is performed to obtain feature F. 01_AB Then, the feature attention module is used to process feature F. 01_AB and feature F 02_SFEM Feature extraction is performed to obtain feature F. 02_FA Then, the feature F is processed through a residual network. 02_FA Feature extraction is performed to obtain feature F. 02_RB ; S5, using convolution on feature F 02_RB Preliminary feature extraction is performed to obtain feature F. 02_AB Then, the feature attention module is used to process feature F. 02_AB and feature F 03_SFEM Feature extraction is performed to obtain feature F. 03_FA Then, the feature F is processed through a residual network. 03_FA Feature extraction is performed to obtain feature F. 03_RB ; S6, using multi-scale fusion modules of two different scales to respectively process feature F 01_RB Feature F 02_RB Feature F 03_RB The feature F is obtained through processing. 01_MF and feature F 02_MF Among them, the scales of the two multi-scale fusion modules are equal to the scales of the non-uniform noise image I1 and image I2, respectively. S7, for feature F 03_RB The feature F is obtained by sequentially processing the data through a residual network and deconvolution. 03_CT ; S8, for feature F 03_CT With feature F 02_MF By stacking along the feature channel dimension, we obtain feature F. cat2_3_1 ; S9, features F are sequentially processed through convolution, residual network, and deconvolution. cat2_3_1 The feature F is obtained through processing. 02_CT ; S10, for feature F 01_MF With feature F 02_CT1 The noise features F of the global image are obtained by stacking them along the feature channel dimension and then sequentially passing them through convolution, residual network, and convolution. noise ; S11, Subtract noise features F from the non-uniform noise image I1 noise The corrected image output is obtained.

2. The infrared non-uniformity correction method according to claim 1, characterized in that: In step S1, the downsampling specifically involves downsampling by a factor of 2 using the nearest neighbor method; In steps S2 and S9, the convolution is a 3×3 convolution with a stride of 1; In step S2, the shallow feature extraction module f SFEM It includes two 3×3 convolutions, three 1×1 convolutions, and a channel concatenation structure. The three 1×1 convolutions are connected in sequence. The output of one 3×3 convolution is connected to the input of the first 1×1 convolution. Another 3×3 convolution is connected between the first and second 1×1 convolutions. The channel concatenation structure is connected between the second and third 1×1 convolutions. The other input of the channel concatenation structure is used to input image I2 or image I3.

3. The infrared non-uniformity correction method according to claim 1 or 2, characterized in that: In steps S3, S4, S5, S7, S9, and S10, the residual network is composed of 8 residual blocks connected in series; in steps S4 and S5, the convolutions are all 2×2 convolutions with a stride of 2; in steps S9 and S10, the convolutions are all 3×3 convolutions with a stride of 1.

4. The infrared non-uniformity correction method according to claim 3, characterized in that: In step S4, the feature attention module includes a matrix dot product structure, a 3×3 convolution, and a matrix summation structure. The 3×3 convolution is connected between the output of the matrix dot product structure and one input of the matrix summation structure. The two inputs of the matrix dot product structure are respectively connected to the output of the convolution and the shallow feature extraction module f corresponding to the non-uniform noise image I2. SFEM The matrix points and the other input end of the structure are connected to the output end of the convolution; In step S5, the feature attention module includes a matrix dot product structure, a 3×3 convolution, and a matrix summation structure. The 3×3 convolution is connected between the output of the matrix dot product structure and one input of the matrix summation structure. The two inputs of the matrix dot product structure are respectively connected to the output of the convolution and the shallow feature extraction module f corresponding to the non-uniform noise image I3. SFEM The matrix points and the other input end of the structure are connected to the output end of the convolution.

5. The infrared non-uniformity correction method according to claim 4, characterized in that: In step S6, the multi-scale fusion module includes three resize modules, one 1×1 convolution, one 3×3 convolution, and one channel concatenation structure. The three inputs of the channel concatenation structure are respectively connected to the three resize modules, and the outputs of the channel concatenation structure are sequentially connected to the 1×1 convolution and the 3×3 convolution. The output of the 3×3 convolution serves as the multi-scale fusion module f. MF The output of the module and the inputs of the three resize modules serve as the three inputs of the multi-scale fusion module, respectively receiving features F. 01_RB Feature F 02_RB and feature F 03_RB .

6. The infrared non-uniformity correction method according to claim 5, characterized in that: In steps S7 and S9, the deconvolution specifically involves performing a double upsampling process through deconvolution.

7. A calibration method based on a lightweight U-MISO network, characterized in that, Includes the following steps: S1, Construct a lightweight U-MISO network, wherein the lightweight U-MISO network is equipped with the infrared non-uniformity correction method according to any one of claims 1 to 6; S2, the lightweight U-MISO network is trained using multiple sets of known non-uniform noise images and corresponding clean infrared images as training sets, resulting in the trained lightweight U-MISO network; S3. Input any non-uniform noise image into the trained lightweight U-MISO network, and use the output of the trained lightweight U-MISO network as the correction result.

Images