High-resolution image fusion method based on one-dimensional sparse aperture rotating pupil

Abstract

The present invention relates to the field of optical interference imaging technology, and in particular to a high-resolution image fusion method based on a one-dimensional sparse aperture rotating pupil. The method comprises: inputting a target image, imaging the target image using a one-dimensional sparse aperture rotating pupil, and obtaining an image for each rotation to an angle; performing a non-subsampled contourlet transform, and obtaining a low-frequency and high-frequency image for each image; performing equal-weighted superposition and normalization on multiple low-frequency images and high-frequency images; repeating the above steps to construct a data set; establishing a dual-branch residual fusion network; extracting high-frequency branch detail information through a residual learning strategy, and performing a global linear transformation on the low-frequency image; fusing the residual map of the high-frequency branch with the adjusted low-frequency image, and outputting a final high-resolution image. The advantages are: accurate extraction of effective information, streamlining of the data set required for training, obtaining full spectrum coverage through fewer sub-apertures, and equivalent imaging effects of a larger aperture.

Description

Technical Field

[0001] The present invention relates to the technical field of optical interference imaging, and in particular to a high-resolution image fusion method based on a one-dimensional sparse aperture rotating pupil. Background Art

[0002] Sparse aperture imaging technology, through multi-subaperture interferometric imaging, can effectively approach the imaging resolution of traditional large-aperture optical systems. However, this technology faces a key technical bottleneck: to achieve wide-range spatial frequency coverage, the system must meet the distribution requirements of different baseline lengths in multiple directions. This will lead to an exponential growth in the number of subapertures, significantly increasing system costs and exacerbating the difficulty of common phase detection. In addition, when the number of subapertures is small, the baseline pairs of the subapertures are also limited. This is especially true for one-dimensional sparse aperture systems, which only have a single baseline direction, resulting in their resolution advantage being limited to that specific direction.

[0003] Through one-dimensional sparse aperture rotating pupil imaging, information from multiple directions can be transmitted with fewer sub-apertures. Image fusion techniques can then complement information from different directions, resulting in high-resolution images. Current multi-image fusion techniques, such as pyramid transform and wavelet transform, have achieved significant success in fields such as medical image fusion and multispectral remote sensing. However, these techniques primarily focus on fusion of features from multiple sources or multi-focus scenarios, and are less adaptable to the anisotropic frequency band distribution characteristic of rotating pupil imaging systems. Summary of the Invention

[0004] In order to solve the above problems, the present invention provides a high-resolution image fusion method based on one-dimensional sparse aperture rotating pupil.

[0005] The present invention aims to provide a high-resolution image fusion method based on one-dimensional sparse aperture rotating pupil, which specifically includes the following steps:

[0006] S1. Input a target image and image the target image using a one-dimensional sparse aperture rotating pupil. Rotate the sparse sub-aperture array of the one-dimensional sparse aperture rotating pupil by 2 j angles, j is a positive integer ≥ 2; after each rotation to an angle, the target is imaged and an image is obtained;

[0007] S2. respectively obtain the 2 j The images are transformed by non-subsampled contourlet, and each image corresponds to a low-frequency image and 2 j High-frequency images, from 2 j Select a high-frequency image that transmits complete high-frequency information from the high-frequency images; j Low-frequency images and filtered 2 jThe high-frequency images are superimposed with equal weights to obtain a superimposed low-frequency image and a superimposed high-frequency image;

[0008] S3. normalizing the superimposed low-frequency image and high-frequency image respectively;

[0009] S4. Repeat steps S1 to S3 for all target images to obtain a set of processed images for each target and construct a dataset for training.

[0010] S5. Establish a dual-branch residual fusion network; extract detailed information of the high-frequency branch through the residual learning strategy, and perform a global linear transformation on the low-frequency image input by the low-frequency branch; fuse the adjusted low-frequency image with the high-frequency residual image of the high-frequency branch to output the final high-resolution image.

[0011] Preferably, the subaperture centers are located on the same straight line and the mirror surfaces do not interfere with each other.

[0012] Preferably, the non-subsampled contourlet transform includes frequency domain pyramid transform and directional filtering; the specific process is as follows:

[0013] S21. Perform a pyramid transform on the acquired images in the frequency domain to decompose each image into a low-frequency image and a high-frequency image;

[0014] S22. Perform directional filtering on the high-frequency image and obtain 2 j A high-frequency image.

[0015] Preferably, step S22 specifically includes: performing directional filtering on the high-frequency image, further decomposing the high-frequency image into high-frequency strips in multiple directions, wherein each high-frequency strip corresponds to high-frequency detail information in a specific direction; obtaining 2 after directional filtering. j Each high-frequency strip corresponds to an image, and the image represents high-frequency detail information in different directions in the spatial domain.

[0016] Preferably, the dual-branch residual fusion network includes two independent feature extraction branches, namely a low-frequency branch and a high-frequency branch, which process the low-frequency information and high-frequency information obtained by S3 respectively;

[0017] The low-frequency branch captures the overall structural information of the image through global average pooling, then uses 1×1 convolution to generate scaling factors and bias parameters, and limits the parameter range through the Tanh activation function. Finally, a global linear transformation is performed on the input low-frequency image.

[0018] The high-frequency branch uses 5×5 convolution to expand the receptive field and extract large-scale detail information; then it refines the details through two 3×3 convolutions and ReLU activation function; and then performs another 3×3 convolution to finally output the high-frequency residual map.

[0019] Preferably, the dual-branch residual fusion network also includes a fusion module, which splices the low-frequency image adjusted by the global linear transformation and the original high-frequency image in the channel dimension, performs preliminary fusion through 5×5 convolution, and then further fuses the information through ReLU activation and 3×3 convolution, and finally generates a spatially adaptive mixing coefficient map through the Sigmoid activation function; the high-frequency residual map and the mixing coefficient map are multiplied element by element, and then superimposed on the adjusted low-frequency image to form the final reconstructed image.

[0020] Preferably, the loss function of the dual-branch residual fusion network is a composite loss function constructed by structural similarity and root mean square error; the low-frequency branch learning rate of the dual-branch residual fusion network is set to 1 -4 , the high-frequency branch learning rate is set to 1 -3 .

[0021] Compared with the prior art, the present invention can achieve the following beneficial effects:

[0022] The present invention provides a high-resolution image fusion method based on a one-dimensional sparse aperture rotating pupil. The non-subsampled contourlet transform (NSCT) is used for one-dimensional sparse aperture rotating pupil imaging to accurately extract effective information from the image, simplify the data, and facilitate subsequent processing. Through fewer sub-apertures, more complete spectrum coverage is obtained, achieving an imaging effect equivalent to a larger aperture. The image quality of the resulting fused image is greatly improved, reaching an image quality far exceeding that of one-dimensional sparse aperture pupil imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of a high-resolution image fusion method based on one-dimensional sparse aperture rotating pupil provided according to an embodiment of the present invention.

[0024] Figure 2 1 is an aperture configuration diagram of a one-dimensional sparse aperture rotating pupil at a specific angle provided by an embodiment of the present invention.

[0025] Figure 3 1 is a schematic diagram illustrating a one-dimensional sparse aperture rotation pupil rotation angle and a decomposition angle of a non-subsampled contourlet transform (NSCT) provided according to an embodiment of the present invention.

[0026] Figure 4 4 is a schematic diagram of a non-subsampled contourlet transform (NSCT) process provided according to an embodiment of the present invention.

[0027] Figure 5 This is an original imaging image of a one-dimensional sparse aperture rotating pupil provided according to an embodiment of the present invention.

[0028] Figure 6The one-dimensional sparse aperture rotating pupil provided by the embodiment of the present invention is Figure 5 Imaging effect diagram.

[0029] Figure 7 1 is a low-frequency image and eight high-frequency images obtained after the non-subsampled contourlet transform provided by an embodiment of the present invention; (A) represents the low-frequency image; (B) represents the eight high-frequency images.

[0030] Figure 8 According to an embodiment of the present invention, Figure 6 Each of the 8 example imaging images in the figure obtains a high-frequency image that conveys complete high-frequency information in the corresponding direction.

[0031] Figure 9 1 and 2 are low-frequency images and high-frequency images obtained by equal-weighted superposition according to an embodiment of the present invention; (A) represents a low-frequency image; (B) represents a high-frequency image.

[0032] Figure 10 2 is a schematic diagram of a dual-branch residual fusion network structure provided according to an embodiment of the present invention.

[0033] Figure 11 is the final image obtained by fusion according to an embodiment of the present invention. DETAILED DESCRIPTION

[0034] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the following description, identical modules are denoted by identical reference numerals. In the case of identical reference numerals, their names and functions are also identical. Therefore, their detailed description will not be repeated.

[0035] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not constitute a limitation of the present invention.

[0036] The present invention provides a high-resolution image fusion method based on one-dimensional sparse aperture rotating pupil. Figure 1 , specifically including the following steps:

[0037] S1. Input the target image and image the target image using a one-dimensional sparse aperture rotating pupil. The one-dimensional sparse aperture rotating pupil includes a sparse sub-aperture array. During imaging, the sparse sub-aperture array is rotated at a certain angle. After each rotation to a certain angle, the target is imaged once to obtain an image. Rotation angle 2 j , j is a positive integer greater than or equal to 2 (j is determined by the properties of NSCT transformation, and the angle of each rotation is in the direction where NSCT can extract details), with the imageA For example, the image at the first angle is .

[0038] The number and diameter D of the subapertures are specified by the user according to specific needs; the centers of the subapertures are located on the same straight line, and there is no interference between the mirror surfaces; during imaging, rotation is performed with the midpoint of the sparse subaperture array as the center of the circle. In a specific embodiment, the rotation angles are -33.75°, -11.25°, 11.25°, 33.75°, -56.25°, -78.75°, 78.75°, and 56.25°. The target is imaged once at the corresponding angle, and the imaging order is not required; the number of subapertures is 3 ( Figure 2 ).

[0039] The method for determining the rotation angle is consistent with the theory of non-subsampled contourlet transform (NSCT), and is set according to the angle selection rule of the directional filtering step of NSCT; the square is divided into 8 equal parts from the center, and all diagonal blocks are regarded as a strip. In this case, there are 4 strips, j=2; when j is greater than 2, on this basis, the outer side length of each strip is divided into 2 j-2 Then connect it to the center of the square, and you will get 2 j The rotation angle is the center position of each strip. Figure 3 As shown, the rotation angle is the dotted line. In the figure, j=3, which corresponds to the center angle of the strip in the NSCT transform. The angles of two adjacent strips of NSCT are different.

[0040] S2. Performing non-subsampled contourlet transform (NSCT) on each of the acquired images to obtain a low-frequency image and several high-frequency images corresponding to each image, and selecting a high-frequency image that conveys complete high-frequency information from the several high-frequency images; superimposing the several low-frequency images and the several high-frequency images with equal weights to obtain a superimposed low-frequency image and a superimposed high-frequency image;

[0041] Specifically, the 2 obtained j images 、 、…… , perform non-subsampled contourlet transform (NSCT), which includes two steps: frequency domain pyramid transform and directional filtering; see Figure 4 The specific process is as follows:

[0042] S21. Perform a pyramid transform on the acquired image in the frequency domain to decompose the image into a low-frequency image and a high-frequency image; the low-frequency image contains the smooth background information of the image, and the high-frequency image contains the detailed information of the image.

[0043] S22. Perform directional filtering on the high-frequency image to further decompose the high-frequency image into high-frequency strips in multiple directions. Each high-frequency strip in each direction corresponds to high-frequency detail information in a specific direction. After directional filtering, 2 j Each high-frequency stripe corresponds to an image, and in the spatial domain, the images represent high-frequency detail information in different directions. It is important to note that the direction of the bands is actually perpendicular to the direction of the texture. For example, imagine two horizontal subapertures interfering with each other. Their modulation transfer function (MTF) shows horizontal stripes, but the interference fringes are vertical.

[0044] S23.From 2 j A high-frequency image that conveys complete high-frequency information is selected from the decomposed high-frequency images.

[0045] Image For example, when performing pyramid transformation, only one transformation is performed, so a low-frequency image is obtained , and 2 j high-frequency images 、 、…… ;lo,hi represent low frequency (low) and high frequency (high) respectively. Because the rotation method of the design is the same as the decomposition method of NSCT, j Among the high-frequency images, only one image contains more spectrum information of the original image, and the other two j -1 high-frequency image does not have this feature. Because the MTF of the one-dimensional sparse aperture rotating pupil rotated to this angle in this application has poor information transmission ability in other directions and cannot transmit high-frequency information in these directions, the final image A 2 j Image of the image A n Each gets a low-frequency image and a high-frequency image ; 2 j low-frequency images and 2 j The high-frequency images are superimposed with equal weights to obtain a low-frequency image and a high-frequency image .

[0046] Brief Principle: This invention uses a one-dimensional sparse aperture with pupil rotation for imaging. Since the MTF in the frequency domain is limited to a single band, meaning only texture information in this direction is transmitted and all other directions are lost, and the MTF cutoff frequency in this direction is very high, a method is needed to extract information in different directions. Combined with aperture rotation, imaging is performed in each direction, extracting the best information in the corresponding direction. This information is then complementary and fused, ultimately achieving fuller spectral coverage with fewer sub-apertures, achieving the equivalent of a larger aperture.

[0047] In a specific embodiment, 8 images are acquired, and after performing non-subsampled contourlet transform, 8 low-frequency images and 64 high-frequency images are obtained. 8 high-frequency images are selected from the 64 high-frequency images; the 8 low-frequency images are equally weighted superimposed to obtain a superimposed low-frequency image; the 8 high-frequency images are equally weighted superimposed to obtain a superimposed high-frequency image;

[0048] The equal-weighted superposition method can effectively combine image information from different angles and provide a basis for subsequent image fusion.

[0049] S3. Normalize the superimposed low-frequency image and high-frequency image separately to obtain a set of processed images, ensuring that the pixel values ​​of the images are within a uniform range; the normalization process can adjust the pixel value range to [0, 1] or other specified range.

[0050] S4. Repeat steps S1 to S3 for all target images, obtain a set of processed images for each target, and construct a dataset for training.

[0051] S5. Establish a dual-branch residual fusion network; extract detailed information of the high-frequency branch through the residual learning strategy, and perform a global linear transformation on the low-frequency image input by the low-frequency branch; fuse the adjusted low-frequency image with the residual image of the high-frequency branch to output the final high-resolution image.

[0052] Specifically, the dual-branch residual fusion network consists of two independent branches, a low-frequency branch and a high-frequency branch, which process the low-frequency information obtained by each image in S3 in the dataset respectively. and high-frequency information ;

[0053] In the low-frequency branch, global average pooling is used to capture the overall structural information of the image. 1×1 convolution is then used to generate scaling factors and bias parameters. The Tanh activation function is then used to limit the parameter range. Finally, a global linear transformation is performed on the input low-frequency image to adjust its overall brightness and contrast.

[0054] The high-frequency branch adopts the strategy of "large convolution kernel + residual learning". The high-frequency branch focuses on learning the residual details of the input image instead of directly outputting the complete features:

[0055] The high-frequency branch uses a 5×5 convolution to expand the receptive field to extract large-scale detail information; then it gradually refines the details through two 3×3 convolutions and a ReLU activation function; and then performs another 3×3 convolution to finally output a high-frequency residual map.

[0056] The fusion module concatenates the low-frequency image adjusted by global linear transformation and the original high-frequency image in the channel dimension, performs preliminary fusion through 5×5 convolution, further fuses the information through ReLU activation and 3×3 convolution, and finally generates a spatial adaptive coefficient map through the Sigmoid activation function to dynamically adjust the weights of information in different frequency bands.

[0057] In the fusion stage, the residual map of the high-frequency branch is modulated according to the spatially adaptive mixing coefficient map. The network multiplies the high-frequency residual map with the mixing coefficient map element by element, and then superimposes it on the adjusted low-frequency image to form the final reconstructed image, that is, the final high-resolution image.

[0058] During the training process, a composite loss function is constructed using structural similarity (SSIM) and root mean square error (MSE) to balance the structural similarity and pixel-level error of the image. The learning rate of the network is set to 1 in the low-frequency branch. -4 , set to 1 in the high frequency branch -3 During training, the batch size is fixed at 1 to accommodate image inputs of varying sizes, and the number of training epochs is set to 300. This allows the network to effectively learn and reconstruct high-quality images while fully utilizing both low-frequency and high-frequency information. After training, the resulting optimized network model is applied to a practical high-resolution image fusion task to produce high-quality fusion results.

[0059] Example 1

[0060] A high-resolution image fusion method based on one-dimensional sparse aperture rotating pupil. The flowchart is shown in Figure 1 , specifically including the following steps:

[0061] S1. Input a target image and image the target image using a one-dimensional sparse aperture rotating pupil. The one-dimensional sparse aperture rotating pupil includes a sparse sub-aperture array with three sub-apertures, whose centers lie on the same straight line and whose mirror surfaces do not interfere with each other.

[0062] like Figure 2As shown in the figure, during imaging, the midpoint of the sparse sub-aperture array is rotated at 8 angles, including -33.75°, -11.25°, 11.25°, 33.75°, -56.25°, -78.75°, 78.75°, and 56.25°. After each rotation to a specific angle, the target is imaged once to obtain an image. The example target image is shown in the figure. Figure 5 As shown, the image obtained by imaging is as follows Figure 6 shown.

[0063] S2. Perform non-subsampled contourlet transform (NSCT) on the 8 images obtained, and obtain a low-frequency image for each image ( Figure 7 Middle A) and 8 high-frequency images ( Figure 7 Middle B), select a high-frequency image that conveys complete high-frequency information from 8 high-frequency images ( Figure 7 The circled picture in Figure B) Figure 6 Each of the eight sample imaging images in the image above obtains a high-frequency image that conveys complete high-frequency information in the corresponding direction (see Figure 8 );

[0064] Perform equal weighted superposition on 8 low-frequency images to obtain a superimposed low-frequency image; perform equal weighted superposition on 8 high-frequency images to obtain a superimposed high-frequency image; the equal weighted superposition method can effectively combine image information from different angles, providing a basis for subsequent image fusion ( Figure 9 ).

[0065] S3. Normalize the superimposed low-frequency image and high-frequency image separately to obtain a set of processed images, ensuring that the pixel values ​​of the images are within a uniform range; the normalization process can adjust the pixel value range to [0, 1] or other specified range.

[0066] S4. Repeat steps S1 to S3 for all target images to obtain a set of processed images for each target and construct a dataset for training.

[0067] S5. Establish a dual-branch residual fusion network; the dual-branch residual fusion network includes two independent feature extraction branches, namely a low-frequency branch and a high-frequency branch, which process the low-frequency and high-frequency information obtained in S3 respectively;

[0068] In the low-frequency branch, global average pooling is used to capture the overall structural information of the image. 1×1 convolution is then used to generate scaling factors and bias parameters. The Tanh activation function is then used to limit the parameter range. Finally, a global linear transformation is performed on the input low-frequency image to adjust its overall brightness and contrast.

[0069] The high-frequency branch adopts the strategy of "large convolution kernel + residual learning". The high-frequency branch focuses on learning the residual details of the input image instead of directly outputting complete features. The high-frequency branch uses 5×5 convolution to expand the receptive field to extract large-scale detail information. It then gradually refines the details through two 3×3 convolutions and ReLU activation functions. Another 3×3 convolution is performed to finally output the residual map.

[0070] The fusion module concatenates the low-frequency image adjusted by the global linear transformation and the original high-frequency image in the channel dimension, performs preliminary fusion through 5×5 convolution, further fuses the information through ReLU activation and 3×3 convolution, and finally generates a spatial adaptive coefficient map through the Sigmoid activation function to dynamically adjust the weights of information in different frequency bands. In the fusion stage, the network multiplies the high-frequency residual map with the mixing coefficient map element by element, and then superimposes it on the adjusted low-frequency image to form the final reconstructed image. See the schematic diagram of the dual-branch residual fusion network structure for details. Figure 10 .

[0071] During the training process, a composite loss function is constructed using structural similarity (SSIM) and root mean square error (MSE) to balance the structural similarity and pixel-level error of the image. The learning rate of the network is set to 1 in the low-frequency branch. -4 , set to 1 in the high frequency branch -3 During training, the batch size is fixed to 1 to adapt to image inputs of different sizes, and the number of training rounds is set to 300. After training, the optimized network model is applied to the actual high-resolution image fusion task to generate high-quality fusion results; the results are shown in Figure 2. Figure 11 shown.

[0072] The eight images obtained by direct imaging and the final fused image are compared with the original images. The peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) results are as follows:

[0073] Table 1 Imaging images and fused images

[0074]

[0075] As shown in the table, the PSNR values ​​of the eight directly imaged images ranged from 23.27dB to 23.45dB, while the PSNR of the final fused image was 28.9122dB, significantly higher than those of the other directly imaged images. This indicates that the fused image has smaller pixel-level errors than the directly imaged images, demonstrating higher image quality. The SSIM values ​​of the directly imaged images ranged from 0.625 to 0.633, while the SSIM value of the final fused image was 0.9525, significantly higher than that of the other images, indicating that the fused image performed better in terms of structural similarity, that is, it was better at preserving image structural information.

[0076] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved. This is not limited herein.

[0077] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A high-resolution image fusion method based on one-dimensional sparse aperture rotating pupil, characterized by: The specific steps include: S1. Input a target image and image the target image using a one-dimensional sparse aperture rotating pupil. Rotate the sparse sub-aperture array of the one-dimensional sparse aperture rotating pupil by 2 j angles, j is a positive integer ≥ 2; after each rotation to an angle, the target is imaged and an image is obtained; S2. respectively obtain the 2 j The images are transformed by non-subsampled contourlet, and each image corresponds to a low-frequency image and 2 j High-frequency images, from 2 j Select a high-frequency image that transmits complete high-frequency information from the high-frequency images; j Low-frequency images and filtered 2 j The high-frequency images are superimposed with equal weights to obtain a superimposed low-frequency image and a superimposed high-frequency image; S3. normalizing the superimposed low-frequency image and high-frequency image respectively; S4. Repeat steps S1 to S3 for all target images to obtain a set of processed images for each target and construct a dataset for training. S5. Establish a dual-branch residual fusion network; extract detailed information of the high-frequency branch through the residual learning strategy, and perform a global linear transformation on the low-frequency image input by the low-frequency branch; fuse the adjusted low-frequency image with the high-frequency residual image of the high-frequency branch to output the final high-resolution image.

2. The high-resolution image fusion method based on one-dimensional sparse aperture rotating pupil according to claim 1, characterized in that: The subaperture centers are located on the same straight line, and the mirror surfaces do not interfere.

3. The high-resolution image fusion method based on one-dimensional sparse aperture rotating pupil according to claim 1, characterized in that: The non-subsampled contourlet transform includes frequency domain pyramid transform and directional filtering; the specific process is as follows: S21. Perform a pyramid transform on the acquired images in the frequency domain to decompose each image into a low-frequency image and a high-frequency image; S22. Perform directional filtering on the high-frequency image and obtain 2 j A high-frequency image.

4. The high-resolution image fusion method based on one-dimensional sparse aperture rotating pupil according to claim 3, characterized in that: The step S22 specifically includes: performing directional filtering on the high-frequency image, further decomposing the high-frequency image into high-frequency strips in multiple directions, where each high-frequency strip corresponds to high-frequency detail information in a specific direction; obtaining 2 j Each high-frequency strip corresponds to an image, and the image represents high-frequency detail information in different directions in the spatial domain.

5. The high-resolution image fusion method based on one-dimensional sparse aperture rotating pupil according to claim 1, characterized in that: The dual-branch residual fusion network includes two independent feature extraction branches, namely a low-frequency branch and a high-frequency branch, which process the low-frequency information and high-frequency information obtained by S3 respectively; The low-frequency branch captures the overall structural information of the image through global average pooling, then uses 1×1 convolution to generate scaling factors and bias parameters, and limits the parameter range through the Tanh activation function. Finally, a global linear transformation is performed on the input low-frequency image. The high-frequency branch uses 5×5 convolution to expand the receptive field and extract large-scale detail information; then it refines the details through two 3×3 convolutions and ReLU activation function; and then performs another 3×3 convolution to finally output the high-frequency residual map.

6. The high-resolution image fusion method based on one-dimensional sparse aperture rotating pupil according to claim 5, characterized in that: The dual-branch residual fusion network also includes a fusion module, which concatenates the low-frequency image adjusted by the global linear transformation and the original high-frequency image in the channel dimension, performs preliminary fusion through 5×5 convolution, further fuses the information through ReLU activation and 3×3 convolution, and finally generates a spatially adaptive mixing coefficient map through the Sigmoid activation function; the high-frequency residual map is element-wise multiplied with the mixing coefficient map, and then superimposed on the adjusted low-frequency image to form the final reconstructed image.

7. The high-resolution image fusion method based on one-dimensional sparse aperture rotating pupil according to claim 5, characterized in that: The loss function of the dual-branch residual fusion network is a composite loss function constructed by structural similarity and root mean square error; the low-frequency branch learning rate of the dual-branch residual fusion network is set to 1 -4 , the high-frequency branch learning rate is set to 1 -3 .

Images