TDI camera dithered degraded image restoration method

Abstract

The application discloses a kind of TDI camera flutter degradation image restoration methods.The present application includes the following steps: first, based on the flutter curve of TDI camera, a one-dimensional equivalent flutter curve is constructed;Then, the one-dimensional equivalent flutter curve is converted into two-dimensional equivalent flutter surface;Then, using two-dimensional equivalent flutter surface, the geometric distortion of the actual shooting TDI degradation image is pre-corrected, and the pre-corrected TDI degradation image is obtained;Again, a simulated TDI degradation image dataset is constructed, and a neural network is trained using the pre-corrected simulated degradation image, to obtain a flutter degradation image restoration network;Finally, using the flutter degradation image restoration network, the actual shooting and pre-corrected TDI degradation image is restored, and the corresponding restored image is obtained.The present application can efficiently and accurately model and simulate the flutter degradation of TDI camera push-broom imaging and image restoration, thereby improving the resolution and imaging quality of remote sensing images.

Description

Technical Field

[0001] This invention relates to a remote sensing image quality improvement method in the field of computational imaging, specifically a method for restoring freak-degraded images from time-delayed and integrated (TDI) cameras. Background Technology

[0002] In the field of high-resolution Earth imaging observations by optical remote sensing satellites, especially in sun-synchronous orbits and very low Earth orbits, TDI cameras are mainly used as imaging sensors. Due to the time-integration characteristics of TDI cameras, their imaging has a higher signal-to-noise ratio, making them suitable for high-resolution imaging under steady-state relative motion and low-light conditions. However, the key to high-quality TDI imaging lies in matching the relative motion between the imaging target and the device with the rate of charge transfer within the TDI detector. Unsteady flutter inevitably exists on the platform, leading to motion mismatch during imaging, causing image blurring and distortion, failing to meet the designed resolution and imaging quality requirements. Post-processing is necessary to further eliminate degradation effects and improve image quality.

[0003] Traditional image restoration methods use point spread functions (PSFs) as prior information to reconstruct sharp images. However, these methods lack consideration for inter-line overlap, misalignment, and content distortion, resulting in drawbacks such as calculating and restoring PSF line by line, high computational cost, and susceptibility to PSF reconstruction errors. In recent years, deep learning has made significant progress in image restoration, demonstrating outstanding effectiveness in addressing image blurring and noise degradation. However, deep learning methods require large amounts of targeted training data, which is relatively difficult to obtain.

[0004] A series of research studies have been conducted both domestically and internationally on flutter degradation simulation and image restoration techniques for low-orbit pushbroom imaging in optical remote sensing. Traditional remote sensing image degradation simulation and restoration methods use blur kernels or analytical modulation transfer functions to process the entire image. For example, Wahballah et al. proposed a method to construct the MTF of the entire image using the pointing accuracy and angular velocity of the TDI camera. However, the blurring of TDI degraded images is usually spatially variable, so the above methods need further improvement. Wang Mi et al. from Wuhan University proposed the CDSM model to model pushbroom imaging. They determined the subdivision number of the CDSM (Continuous Dynamic Shooting Model) model through an integral transform function to achieve more accurate PSF estimation and combined it with Wiener filtering for context-based adaptive image restoration. Liu Shijie et al. from Tongji University further improved the PSF compensation based on this, enhancing the quality of the restored image. Researchers at Jilin University proposed a vibration modulation transfer function based on statistical properties, applying different vibration MTFs to each row to simulate spatially variable degraded images. Researchers from Harbin Institute of Technology and Zhejiang University proposed a line-by-line PSF convolution process. For each time series, the PSF at that moment needs to be estimated first using the relative motion velocity. Then, the blurred lines of different time series are superimposed to simulate the charge accumulation process. This can simulate the image degradation process caused by flutter relatively accurately. However, when the relative motion is complex, it is difficult to analytically calculate the degraded PSF. The image degradation simulation methods mentioned above are all based on PSF convolution, which lacks consideration for inter-line superposition, misalignment, and content deformation. They suffer from the disadvantages of calculating PSF line by line and restoring line by line, resulting in high computational cost. Furthermore, the restoration results are easily affected by PSF reconstruction errors.

[0005] With the rapid development of artificial intelligence, many research teams are dedicated to using deep learning to solve image restoration problems. For example, Syed Waqas Zamir et al. from the Origin AI Research Institute proposed Restormer, which uses a multi-head attention mechanism and a feedforward neural network to achieve long-distance pixel interaction and the ability to process large-size images. Sun Jian et al. from Kuangshi proposed NAFNet, a network without activation functions, based on Unet, introducing LayerNorm, simplifying gating units, and simplifying channel attention mechanisms. NAFNet abandons redundant attention and activation function designs and has the characteristics of fast training convergence speed, small number of network parameters, and low computational cost. Although learning-based image restoration methods have achieved good results, these methods are often used for restoration with area scan cameras, and the degradation factors they consider are mainly blurring caused by camera shake, rather than degradation factors of TDI cameras. Summary of the Invention

[0006] To address the problems and needs existing in the background technology, this invention provides a method for simulating and restoring TDI camera flutter degradation. This invention solves problems such as rapid simulation of pushbroom imaging, dataset construction, and high-quality image restoration through pushbroom imaging flutter degradation modeling and simulation, image geometric distortion pre-correction based on flutter measurement data, and high-efficiency fine image restoration based on lightweight neural networks, further improving the quality of low-orbit pushbroom imaging. The in-depth research and modeling simulation of the pushbroom imaging flutter degradation mechanism in this invention will help to more accurately restore degraded images, and also provides a data and methodological foundation for the application of deep learning in the field of remote sensing.

[0007] The technical solution of the present invention is as follows:

[0008] I. A method for restoring images degraded by flutter in TDI cameras

[0009] (1) Construct a one-dimensional equivalent flutter curve based on the flutter curve of a TDI camera;

[0010] (2) Transform the one-dimensional equivalent flutter curve into a two-dimensional equivalent flutter surface;

[0011] (3) Use the two-dimensional equivalent flutter curve to perform geometric distortion pre-correction on the actual captured TDI degradation image to obtain the pre-corrected actual TDI degradation image;

[0012] (4) Train a neural network using the pre-corrected TDI degradation simulation image dataset to obtain a flutter degradation image restoration network;

[0013] (5) Use the flutter degradation image restoration network to restore the pre-corrected actual TDI degradation image to obtain the corresponding restored image.

[0014] In (1), for the jitter curves in the x or y direction, the jitter curves of different levels at each pixel position are accumulated and sampled during each exposure time to obtain a one-dimensional equivalent jitter curve describing the current pixel position offset during the entire push-broom imaging process.

[0015] Specifically, (2) refers to:

[0016] After combining the one-dimensional equivalent dizziness curves in the x and y directions, the pixel dizziness state is then copied and extended along the direction perpendicular to the push-broom to obtain the equivalent dizziness surface that describes the positional offset of each pixel in the two-dimensional push-broom image.

[0017] In step (3), the negative value of the equivalent flutter surface is used as the flutter correction optical flow, and the flutter correction optical flow is used to perform geometric distortion pre-correction on the actual captured TDI degraded image.

[0018] In step (4), the steps for constructing the pre-corrected TDI degraded image dataset include:

[0019] For the aerial image dataset, firstly, each image in the aerial image dataset is converted into a grayscale image and then converted to the luminance domain to obtain a luminance domain image; then, flutter degradation simulation is performed on each luminance domain image, followed by geometric distortion pre-correction to obtain a pre-corrected TDI degradation simulation image and the corresponding equivalent flutter surface. All images are processed in turn to obtain a pre-corrected TDI degradation simulation image dataset.

[0020] After performing flutter degradation simulation and pre-correction on each luminance domain image, the pre-corrected TDI degradation simulation image and the corresponding equivalent flutter surface are obtained, specifically:

[0021] First, flutter curves in the x and y directions are randomly generated and denoted as simulated flutter curves. Then, a corresponding one-dimensional equivalent flutter curve is constructed based on the simulated flutter curve. Next, the one-dimensional equivalent flutter curve is converted into a two-dimensional equivalent flutter surface. Then, each brightness domain image is resampled to obtain a flutter degradation simulation image. The two-dimensional equivalent flutter surface is used to perform geometric distortion pre-correction on the flutter degradation simulation image to obtain a pre-corrected TDI degradation simulation image.

[0022] In (4) and (5), the input to the flutter degradation image restoration network is a pre-corrected TDI degradation simulation image and a pre-corrected actual TDI degradation image, or a pre-corrected TDI degradation simulation image and an equivalent flutter surface, and a pre-corrected actual TDI degradation image and an equivalent flutter surface.

[0023] II. A computer device

[0024] The device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method.

[0025] III. A computer-readable storage medium

[0026] The medium stores a computer program that, when executed by a processor, implements the steps of the method.

[0027] IV. A computer program product

[0028] The product includes a computer program / instructions that, when executed by a processor, implement the steps of the method.

[0029] The beneficial effects of this invention are as follows:

[0030] This invention can efficiently and accurately model and simulate the jitter degradation of push-broom imaging by TDI cameras, providing a large amount of targeted training data for deep learning methods;

[0031] This invention combines distortion pre-correction with neural network fine restoration, which can effectively improve the resolution and imaging quality of remote sensing images. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the simulation and image restoration technology process for TDI camera flutter degradation.

[0033] Figure 2 This is a schematic diagram of pushbroom imaging using a TDI camera.

[0034] Figure 3 The calculation process for the equivalent flutter curve is shown, where (a) is the flutter curve and (b) is the equivalent flutter model.

[0035] Figure 4 The equivalent chatter curves are robust to measurement noise. (a) is the ideal chatter curve, (b) is the chatter curve with noise, (c) is the equivalent chatter curve of (a), and (d) is the equivalent chatter curve of (b).

[0036] Figure 5 This is a simulation process for the equivalent flutter surface and flutter degradation.

[0037] Figure 6 This is a schematic diagram of location resampling.

[0038] Figure 7 The images are simulated degradation images with different TDI integration levels, where (a) is the original image, (b) is the simulated degradation image with integration level k=1, (c) is the simulated degradation image with integration level k=16, and (d) is the simulated degradation image with integration level k=128.

[0039] Figure 8 The images are simulated degradation images with different flutter amplitudes, where (a) is a simulated degradation image with a flutter amplitude of 1 pixel, (b) is a simulated degradation image with a flutter amplitude of 7 pixels, (c) is a simulated degradation image with a flutter amplitude of 12 pixels, and (d) is a simulated degradation image with a flutter amplitude of 16 pixels.

[0040] Figure 9 The images are simulated degradation images at different flutter frequencies, where (a) is a simulated degradation image at a flutter frequency of 30 pixels, (b) is a simulated degradation image at a flutter frequency of 100 pixels, (c) is a simulated degradation image at a flutter frequency of 300 pixels, and (d) is a simulated degradation image at a flutter frequency of 1000 pixels.

[0041] Figure 10The diagram illustrates the pre-correction of distortion and fine image restoration, where (a) is a flowchart of the pre-correction of distortion and fine image restoration process, (b) is a flowchart of the pre-correction of distortion and fine image restoration process, and (c) is a diagram of the neural network structure in fine image restoration (the input is the degraded image).

[0042] Figure 11 This is another fine image restoration network structure diagram (the input is an equivalent flutter surface and a degraded image, and the geometric distortion pre-correction stage is integrated into the image restoration network).

[0043] Figure 12 The images shown are the restoration results of low-frequency flutter distortion degradation images, where (a) is the reference image, (b) is the low-frequency flutter distortion degradation image, (c) is the restoration result of the MIRNetV2 method, and (d) is the restoration result of the present invention.

[0044] Figure 13 The images shown are the restoration results of high-frequency flutter blurring degradation images, where (a) is the reference image, (b) is the high-frequency flutter distortion degradation image, (c) is the restoration result of the MIRNetV2 method, and (d) is the restoration result of the present invention. Detailed Implementation

[0045] To enable those skilled in the art to better understand the present invention and to more clearly define the scope of protection of the present invention, the present invention will be described in detail below with reference to certain specific embodiments and accompanying drawings. It should be noted that the following are only some specific embodiments of the present invention, and are merely a part of the embodiments of the present invention. The specific and direct descriptions of related structures are only for the convenience of understanding the present invention, and the specific features do not necessarily or directly limit the scope of the present invention. Conventional selections and substitutions made by those skilled in the art under the guidance of the present invention, as well as reasonable arrangements and combinations of several technical features under the guidance of the present invention, should all be considered within the scope of protection of the present invention.

[0046] To address the image quality degradation problem caused by flutter in TDI cameras and improve the pushbroom imaging quality of TDI cameras, this invention discloses a method for restoring images degraded by flutter in TDI cameras, such as... Figure 1 As shown, it comprises three parts: "TDI flutter degradation modeling," "TDI flutter degradation image simulation," and "TDI flutter degradation image restoration." The specific implementation includes the following steps:

[0047] (1) Construct a one-dimensional equivalent flutter curve based on the flutter curve of a TDI camera;

[0048] For a TDI camera that includes a gyroscope (or other pose measurement device), the angular velocity ω at time t is measured by the gyroscope about the x-axis. x(t), the angular position θ of the optical axis relative to the x-direction at the initial moment. x (0) = 0, the exposure time is T, and the position of the optical axis swing angle at time t (t1≤t≤t1+T) is:

[0049]

[0050] Where, ω x (t′) represents the angular velocity at time t′.

[0051] set up Where, θ x (t) represents the angular position of the optical axis relative to the x-direction at time t, A i f is the amplitude. i For frequency, For phase, such as Figure 2 As shown, the object-image relationship during the imaging process, and the relationship between the measured jitter curve and the displacement of the pixel plane on the sensor are as follows:

[0052]

[0053] Among them, J x (t) represents the flutter curve in the x-direction, l represents the image distance, and P represents the pixel size.

[0054] For a TDI camera with an integration level of K, the timing start of recording differs across levels, but they share a common jitter curve. To leverage this characteristic, an equivalent jitter curve can be calculated. For the jitter curves in the x or y direction, the jitter curves of different levels at each pixel location are accumulated and sampled within each exposure time level to obtain a one-dimensional equivalent jitter curve describing the current pixel position shift throughout the pushbroom imaging process, such as... Figure 3 This is a schematic diagram illustrating the calculation of the flutter curve in the x-direction. The formula for calculating the one-dimensional equivalent flutter curve is:

[0055]

[0056] In the formula, X(t) and Y(t) are the equivalent flutter curves in the x-direction and y-direction, respectively, and J x The flutter curve is in the x-direction, J y Let t represent the flutter curve in the y-direction, K represent the integration series, M represent the number of flutter curve sampling subdivisions within each exposure time, τ represent the single exposure time, k represent the series number, and m represent the sampling number.

[0057] The equivalent chatter curve can mitigate the effects of chatter measurement noise. For example... Figure 4 As shown, under the condition of measurement noise, the noisy equivalent flutter curve is consistent with the ideal equivalent flutter curve, and the noise does not have a significant impact on the equivalent flutter curve.

[0058] (2) Transform the one-dimensional equivalent flutter curve into a two-dimensional equivalent flutter surface;

[0059] (2) Specifically:

[0060] Since linear arrays of the same level exhibit the same vibration state in the sweep direction during the push-broom process, and each pixel in the linear imaging array exhibits the same jitter state, the state of each pixel position can be obtained by copying the jitter curve. After combining the one-dimensional equivalent jitter curves in the x and y directions, the pixel jitter state is then copied and extended along the direction perpendicular to the push-broom, based on the combined pixel jitter state, resulting in an equivalent jitter surface describing the offset of each pixel position in the two-dimensional push-broom image. In this embodiment, k levels of equivalent jitter curves in the x and y directions are combined and copied h times, where h is the number of pixels perpendicular to the push-broom direction. The resulting equivalent jitter surface F(x,y) describing the offset of each pixel position (x,y) in the two-dimensional push-broom image is:

[0061] F(x,y)=concat{repeat[X(t),h],repeat[Y(t),h]}

[0062] Here, `concat` represents concatenating the two-dimensional tensor along its third dimension, and `repeat[X(t),h]` represents copying the one-dimensional curve X(t) h times along the vertical sweep direction. X(t) and Y(t) are the equivalent dithering curves in the x and y directions, respectively, and t is a positive integer. The number of pixels parallel to the sweep direction is determined by the sweep duration. To facilitate image processing, the image can be cropped into a rectangle along the sweep direction, and the equivalent dithering surface is also cropped to the corresponding size. Figure 5 A schematic diagram containing the equivalent flutter surface.

[0063] (3) The negative value of the equivalent flutter surface is used as the flutter correction optical flow. The flutter correction optical flow is used to perform geometric distortion pre-correction on the actual captured TDI degradation image to obtain the pre-corrected actual TDI degradation image.

[0064] (4) Train a neural network using the pre-corrected TDI degradation simulation image dataset to obtain a flutter degradation image restoration network;

[0065] In (4), the steps for constructing the pre-corrected TDI degradation simulation image dataset include:

[0066] For high-quality aerial image datasets, firstly, each high-resolution RGB aerial image in the dataset is converted to grayscale, then inverse gamma correction is applied, and the image is converted to the luminance domain to obtain a luminance domain image. Next, flutter degradation simulation and geometric distortion pre-correction are performed on each luminance domain image to obtain a pre-corrected TDI degradation simulation image and the corresponding equivalent flutter surface, such as... Figure 5 As shown, all aerial RGB images are processed and used as ground truth images to obtain a pre-corrected TDI degradation simulation image dataset.

[0067] After performing flutter degradation simulation and pre-correction on each luminance domain image, the pre-corrected TDI degradation simulation image and the corresponding equivalent flutter surface are obtained, specifically:

[0068] First, flutter curves in the x and y directions are randomly generated to simulate the flutter state of a TDI camera, denoted as simulated flutter curves. Then, a corresponding one-dimensional equivalent flutter curve is constructed based on the simulated flutter curve. Next, the one-dimensional equivalent flutter curve is converted into a two-dimensional equivalent flutter surface. Then, each brightness domain image is resampled to obtain a flutter degradation simulation image. The two-dimensional equivalent flutter surface is used to perform geometric distortion pre-correction on the flutter degradation simulation image to obtain a pre-corrected TDI degradation image.

[0069] The specific mathematical representation is as follows: the flutter of each level of the linear array is considered independently, and the equivalent flutter curves of the k-th level in the x and y directions of the fine sampling are considered. k (t), y k (t) is:

[0070]

[0071] Where τ is the exposure time per burst. The x and y direction curves are combined to form the k-th level two-dimensional jitter degradation information f, which describes the pixel position offset. k (x,y):

[0072] f k (x,y)=concat{repeat[x k (t),h],repeat[y k (t),h]}

[0073] h is the number of pixels in the TDI linear array (perpendicular to the push-broom direction);

[0074] The k-th level degraded image is:

[0075] B k =grid_sample(I,f k )+n k

[0076] Among them, B k It is a fuzzy graph of a single-level integral, n k It's noise, I is the brightness domain image. grid_sample is position resampling, meaning it's based on the two-dimensional flutter offset information f k Resample the pixel locations of the clear image I. Figure 6 This is a schematic diagram of the location resampling process. Black dots represent pixel locations, and squares represent pixels.

[0077] The intensity value of the original pixel is transferred to the sampled pixel location through pixel resampling. The TDI pushbroom image is an accumulation of each stage, and the final degraded image is as follows:

[0078]

[0079] Based on the above formula, image degradation caused by factors such as imaging platform jitter amplitude, frequency, TDI level, exposure time, and noise can be simulated. Figure 7 of (a), Figure 7 (b) Figure 7 (c) and Figure 7 (d) Figure 8 of (a), Figure 8 (b) Figure 8 (c) and Figure 8 (d) and Figure 9 of (a), Figure 9 (b) Figure 9 (c) and Figure 9 (d) represents the simulation results of flutter degradation images under different parameter settings.

[0080] Figure 6 , Figure 10 (a) and Figure 10 (b) illustrates this process, where the distortion of the image is initially repaired with the help of dithering correction optical flow, but the details of the image still suffer from burrs or blurring.

[0081] (5) Based on the pre-corrected actual TDI degradation image, the image is restored using the flutter degradation image restoration network to obtain the corresponding restored image.

[0082] In (4) and (5), the input to the flutter degradation image restoration network is the pre-corrected actual TDI degradation image, or the pre-corrected actual TDI degradation image and the equivalent flutter surface.

[0083] When the input to the flutter degradation image restoration network is only a pre-corrected TDI degradation image, the available network structures are as follows: Figure 10 As shown in (c).

[0084] When the input to the flutter degradation image restoration network is a pre-corrected TDI degradation image and an equivalent flutter surface, the geometric distortion pre-correction stage is integrated into the image restoration network. Available network structures include... Figure 11 As shown.

[0085] For any image restoration neural network, the simulated, pre-corrected TDI-degraded image B' and the equivalent flutter surface F are taken as input, and the clear, undegraded image I is taken as the target. The restoration parameter is θ, and the output of the restoration network is... Network() is an image restoration neural network. The optimization problem can be expressed as:

[0086]

[0087] in Let be the loss function. By minimizing the above loss function, the parameters of the restoration network are optimized so that it can generate a restoration result on the input degraded image that is as close as possible to the target clear image I.

[0088] The runtime of the method of this invention was tested on a single TDI degradation image with a size of 1024x1024. Table 1 shows the results on a single RTX 3090 GPU, with the results averaged across 10 images. The test was conducted on... Figure 10 The method is demonstrated in Stage 1, which represents the pre-correction stage, and Stage 2, which represents the fine restoration stage. It can be seen that the pre-correction stage takes approximately 0.13 seconds, while the total time is around 1 second, proving the efficiency of the method proposed in this invention.

[0089] Table 1 shows the statistics of the two-stage operation time.

[0090] Runtime (in seconds) 0.13 1.05 1.18

[0091] This invention uses PSNR and SSIM parameters to measure the restoration effect of the method on simulated flutter degradation images. Tables 2 and 3 represent the restoration results for TDI level 1 and TDI level 128, respectively. It can be seen that the method of this invention can achieve high objective parameters in both cases. Figure 12 (a)- Figure 12 (d) and Figure 13 (a)- Figure 13 (d) are all visualized restoration results, showing that the restoration results obtained by the method proposed in this invention have richer details. The above results also prove the effectiveness of the method of this invention.

[0092] Table 2 shows the restoration parameters for flutter degradation images with a TDI level of 1.

[0093] Degraded image 13.57 0.0814 Restored image 39.02 0.9493

[0094] Table 3 shows the restoration parameters for flutter degradation images with a TDI level of 128.

[0095] Degraded image 13.66 0.0904 Restored image 36.52 0.9072

[0096] Finally, it should be noted that the above embodiments and descriptions are only used to illustrate the technical solutions of the present invention and not to limit it. 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 disclosure of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for restoring images degraded by fretting in a TDI camera, characterized in that, Includes the following steps: (1) Construct a one-dimensional equivalent flutter curve based on the flutter curve of a TDI camera; In (1), for the jitter curve in the x or y direction, the jitter curves of different levels at each pixel position are accumulated and sampled during each exposure time to obtain a one-dimensional equivalent jitter curve describing the current pixel position offset during the entire push-broom imaging process. The formula for calculating the one-dimensional equivalent flutter curve is: in, and These are the equivalent flutter curves in the x-direction and the y-direction, respectively. The flutter curve is shown in the x-direction. Let t represent time, K be the integration series, M be the number of flutter curve samples per exposure time, and τ be the exposure time. The index is the sequence number of the series. The sampling sequence number; (2) Transform the one-dimensional equivalent flutter curve into a two-dimensional equivalent flutter surface; Specifically, (2) refers to: After combining the one-dimensional equivalent dizziness curves in the x and y directions, the pixel dizziness state is then copied and extended along the direction perpendicular to the push-broom to obtain the equivalent dizziness surface that describes the positional offset of each pixel in the two-dimensional push-broom image. The two-dimensional equivalent flutter surface satisfies the following formula: in, For a two-dimensional equivalent flutter surface, concat represents the concatenation of the two-dimensional tensor in the third dimension, and repeat... Represents a one-dimensional curve Copy h times in the vertical push-broom direction; (3) Use the two-dimensional equivalent flutter curve to perform geometric distortion pre-correction on the actual captured TDI degradation image to obtain the pre-corrected actual TDI degradation image; (4) Train a neural network using the pre-corrected TDI degradation simulation image dataset to obtain a flutter degradation image restoration network; In step (4), the steps for constructing the pre-corrected TDI degraded image dataset include: For the aerial image dataset, firstly, each image in the aerial image dataset is converted into a grayscale image and then converted to the luminance domain to obtain a luminance domain image; then, flutter degradation simulation is performed on each luminance domain image, followed by geometric distortion pre-correction to obtain a pre-corrected TDI degradation simulation image and the corresponding equivalent flutter surface. All images are processed in turn to obtain a pre-corrected TDI degradation simulation image dataset. After performing flutter degradation simulation and pre-correction on each luminance domain image, the pre-corrected TDI degradation simulation image and the corresponding equivalent flutter surface are obtained, specifically: First, flutter curves in the x and y directions are randomly generated and denoted as simulated flutter curves. Then, a corresponding one-dimensional equivalent flutter curve is constructed based on the simulated flutter curve. Next, the one-dimensional equivalent flutter curve is converted into a two-dimensional equivalent flutter surface. Then, each brightness domain image is resampled to obtain a flutter degradation simulation image. The two-dimensional equivalent flutter surface is used to perform geometric distortion pre-correction on the flutter degradation simulation image to obtain a pre-corrected TDI degradation simulation image. (5) Use the flutter degradation image restoration network to restore the pre-corrected actual TDI degradation image to obtain the corresponding restored image.

2. The method for restoring images degraded by fretting in a TDI camera according to claim 1, characterized in that, In step (3), the negative value of the equivalent flutter surface is used as the flutter correction optical flow, and the flutter correction optical flow is used to perform geometric distortion pre-correction on the actual captured TDI degraded image.

3. The method for restoring images degraded by fretting in a TDI camera according to claim 1, characterized in that, In (4) and (5), the input to the flutter degradation image restoration network is a pre-corrected TDI degradation simulation image and a pre-corrected actual TDI degradation image, or a pre-corrected TDI degradation simulation image and an equivalent flutter surface, and a pre-corrected actual TDI degradation image and an equivalent flutter surface.

4. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 3.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

6. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1 to 3.

Images