A spacecraft image shadow removing method based on multi-illumination angle image fusion

Abstract

The application discloses a spacecraft image shadow removing method based on multi-illumination angle image fusion, uses three-dimensional animation rendering and manufacturing software to manufacture a synthesized spacecraft image, centrally uses a spacecraft model, the spacecraft model is collected from the Internet, the collected model includes different document formats, firstly, all 3D document formats are converted into fbx format, unreal engine generates a simulation image similar to a real scene, too much illumination angle image fusion, a method architecture for reconstructing a shadow-free image from two shadowed spacecraft images, task problem analysis and overall method framework explanation, according to the low pixel brightness characteristics of the shadow area, the input image {I1, I2} is converted into a gray image, the pixel value is compared with the threshold value one by one, the encoder-decoder network is pre-trained in a large natural image data set, and the loss function and training steps in TransMEF and other implementation details are followed, and the verification scene in the MIAS data set is used to evaluate the shadow removing result.

Description

Technical Field

[0001] This invention relates to the field of spacecraft image shadow removal technology, specifically a method for spacecraft image shadow removal based on multi-light angle image fusion. Background Technology

[0002] On-orbit services such as docking, berthing, refueling, repair, rescue, and on-orbit debris removal are considered the most promising ways to advance spacecraft lift time and reduce space debris. Before conducting on-orbit operations, one of the most important tasks is to identify spacecraft components to obtain target locations.

[0003] Spacecraft component detection methods based on 3D LiDAR data are reliable and robust. However, 3D LiDAR systems are costly and heavy, and the large volume of 3D point cloud data results in low computational efficiency, making airborne processing difficult. Space cameras offer an economical alternative with smaller mass, power, and geometric constraints, and their image data can be easily processed onboard without signal delay and significant manpower costs. Therefore, spacecraft component recognition based on spacecraft images has received increasing attention in recent years, especially after the great success of deep learning methods in image recognition. Xu et al. compared the performance of different deep learning-based object detection methods, Faster R-CNN and YOLOv3, in solving the problem of solar panel detection on non-cooperative spacecraft. Feng et al. proposed an unsupervised spacecraft image foreground extraction algorithm that can robustly extract the spacecraft body from the image. Qiu et al. proposed a method derived from 3D LiDAR data. The Max and STK software's spacecraft identification dataset contains 4,500 images of 23 satellites. However, this dataset is synthetic and can only be used to detect solar panels. To further advance the development of spacecraft component identification tasks, Dung et al. collected 3,117 satellite images from synthetic or real images, labeled them through self-supervised learning, and tested several state-of-the-art object detection and semantic segmentation algorithms on the dataset.

[0004] However, a crucial issue has been overlooked in these studies: compared to LiDAR, space cameras are severely limited by insufficient space lighting conditions. In the space environment, the light distribution on the spacecraft surface is primarily affected by direct sunlight, while the diffuse reflection effect from Earth is weak. Therefore, mutual occlusion between different parts of the spacecraft results in shadows with poor visibility in the images. These shadows lead to missing information in the spacecraft images, thereby reducing the accuracy of spacecraft component identification. Therefore, an improved technology is urgently needed to address this problem in existing technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a method for removing shadows from spacecraft images based on multi-light angle image fusion, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for removing shadows from spacecraft images based on multi-illumination angle image fusion, characterized in that: the...

[0007] Data collection: Using 3D animation rendering and production software to create composite images of spacecraft;

[0008] Spacecraft Model Acquisition: The dataset uses spacecraft models, which are collected from the Internet. The collected models include different document formats. First, all 3D document formats must be converted to FBX format.

[0009] Rendering process: Unreal Engine generates simulated images that resemble the real scene;

[0010] Methodology Architecture: This paper presents a methodology for reconstructing a shadow-free image from two shadowed spacecraft images through multi-light angle image fusion, and provides an analysis of the mission problem and an explanation of the overall methodology framework.

[0011] Shadow detection: Based on the characteristic that pixels in shadow areas have low brightness, the input image {I 1, I2} converts the grayscale image and compares the pixel values ​​one by one with the threshold.

[0012] Image feature extraction and reconstruction: After degrading the image quality, reconstruction is performed to improve the robustness of the encoder and decoder in the face of image noise;

[0013] Weight calculation: The loss function is designed by combining supervised learning and self-supervised learning.

[0014] Experimental Results and Analysis: Encoder-decoder networks were pre-trained on a large natural image dataset.

[0015] Preferably, the data collection includes spacecraft model acquisition and rendering processes;

[0016] The process of acquiring and saving images of the spacecraft model involves the following steps: 3D spacecraft model assembly, 3D model format conversion, initial rendering software settings, lighting angle adjustment, rendering of the 3D model, and saving of the spacecraft image. Preferably, after importing the spacecraft 3D model into Unreal Engine 4, atmospheric fog, sky sphere, skylight, and spherical reflection capture are disabled, and the light source is set to directional light.

[0017] Preferably, during the rendering process, after importing the spacecraft 3D model into Unreal Engine, atmospheric fog, sky sphere, skylight, and spherical reflection capture are turned off, and the light source is set to directional light;

[0018] Adjust the angle of the directional light to obtain spacecraft images under multiple illumination angles, and name the dataset "Spacecraft Image Set under Multiple Illumination Angles";

[0019] The dataset contains 50 training scenarios and 17 validation scenarios. Each scenario consists of two sequences of spacecraft images from different lighting angles, ground-based images of the spacecraft under uniform lighting conditions, and semantic segmentation masks for labeled spacecraft components.

[0020] Preferably, the overall method framework in the method architecture requires two spacecraft images under different lighting angles as input, and it is assumed that these two images have been configured to the same camera viewpoint using an image registration method, where I1∈R. H ×W×3 and I2∈R H×W×3 This represents two images of a spacecraft under shadow at different lighting angles, I f ∈R H×W×3 To obtain a fused spacecraft image free of shadow defects, the spacecraft image fusion task can be described as follows:

[0021] I f =φ(I 1, I2) (1)

[0022] In the formula, φ(·) represents the paired shadow spacecraft image {I 1, Our image fusion method reconstructs spacecraft images from I2} to obtain shadow-free images. The shadow detection module of our method architecture extracts spacecraft image shadows (excluding those against the cosmic background) by comparing similarities and differences between spacecraft images at different lighting angles. The extracted spacecraft image shadow region mask can be represented as {S}. 1, S2} Therefore, by performing a pixel-level product between the input image and the shadow mask as shown in Equation (2), we are able to obtain the information complementary image {C} of each input image. 1, C2}

[0023]

[0024] The input shadowed spacecraft image and the compensated image are converted into the YCbCr color space. The luminance channel Y is fused by the proposed network, and the chrominance (Cb and Cr) channels are weighted and fused. The fused luminance and chrominance channels are then reconstructed into an RGB image.

[0025] Preferably, the shadow detection threshold is set to 50 based on the test results. Pixel areas with values ​​less than the threshold are considered shadow areas, which are assigned a value of 1, while the rest are assigned a value of 0.

[0026] The initial shadow mask obtained is {S1,S2}. In the above steps, the algorithm compares two images. Their cosmic background is consistent, but the shadow areas are inconsistent. The algorithm compares the differences between the two initial shadow masks pixel by pixel. The same area is considered as the cosmic background and is assigned a value of 0. The rest remain unchanged.

[0027] Obtain the shadow mask after removing the background. Finally, according to formula (3), the complementary information image corresponding to each input image is obtained;

[0028] In the process of shadow detection algorithm, the input image {I1,I2} is converted into a grayscale image, and the pixel values ​​are compared with the threshold (50) one by one. The pixel areas that are less than the threshold are considered as shadow areas. The shadow areas are assigned a value of 1, and the rest are 0. The initial shadow mask obtained is {S1,S2}.

[0029] The two initial shadow masks are compared pixel by pixel. Identical areas are considered as the cosmic background and assigned a value of 0, while the rest remain unchanged, resulting in the shadow mask after removing the background. Input image and shadow mask pair Perform pixel-by-pixel multiplication to obtain the complementary information image corresponding to the input image.

[0030] Preferably, the Y-channel images of the input image and the complementary information image are input into the encoder network to extract features, and the output of the feature extraction is a feature map. Simultaneously, the input image and its corresponding complementary image are jointly input into the third module—the weight calculation module. Preferably, the weight calculation process can be expressed as:

[0031] W i =H(I i C i ),i∈{1,2} (3)

[0032] H(·) represents the weight calculation function, which is fitted using a three-layer neural network, W i Let represent the fusion weights used to fuse the feature maps of the i-th image and its complementary image. The fusion process can be represented by equation (4). After fusing the two input images with their complementary images respectively, the feature maps are obtained. Figure 1 and characteristics Figure 2 Element-level addition and fusion are performed to obtain the final feature map F. fuse It is reconstructed into an image by the decoder.

[0033]

[0034] Preferably, based on the characteristic of low pixel brightness in the shadow area, the input image {I1,I2} is first converted into a grayscale image, and the pixel values ​​are compared with the threshold value one by one. According to the test results, the threshold is set to 50. Pixel areas with brightness less than the threshold are considered as shadow areas. The shadow areas are assigned a value of 1, and the rest are assigned a value of 0. The resulting initial shadow mask is {S1,S2}.

[0035] The two initial shadow masks are compared pixel by pixel. Identical areas are considered as the cosmic background and assigned a value of 0, while the rest remain unchanged, resulting in the shadow mask after removing the background. Finally, according to formula (3), the complementary information image corresponding to each input image is obtained.

[0036] Preferably, the loss function is designed by combining supervised learning and self-supervised learning in the weight calculation. The input image pairs {I1,C1} and {I2,C2} can obtain fusion weights W1 and W2 through the weight calculation network. After the fusion process of the above equation (7), the fusion feature F can be obtained. fuse_1 F fuse_2 and F fuse These three fused features are all reconstructed by the decoder to obtain the output image I. out1 I out2 and I out First, the image similarity loss between the two images is calculated. This is a self-supervised loss that can be used to constrain the brightness distribution of shadow and non-shadow regions to be consistent. Then, the image similarity loss between the two images and the ground truth image with good lighting is calculated to constrain the global image quality after weight fusion. I is then calculated. out True value of the same well-lit image I G Image similarity loss; structural similarity loss measures the structural similarity between the input image and the reconstructed image, helping the network to better preserve image structural information. Its calculation formula is:

[0037] L ssim =1-SSIM(I in ,I out (5)

[0038] The network loss function for weighted calculation can be expressed as:

[0039] Loss = L ssim (I out1, I out2 )+L ssim (I out ,I G (6)

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] This method creates publicly available spacecraft image datasets under multiple lighting angles and proposes a novel image fusion framework for shadow removal. Unlike handcrafted image fusion strategies, it designs a weighted sub-network block to learn the optimal feature fusion strategy to reduce the differences in lighting intensity distribution in the input spacecraft images, thereby removing image shadows. It combines supervised loss and self-supervised loss functions to constrain the consistency of brightness distribution in shadow and non-shadow areas. Attached Figure Description

[0042] Figure 1 This is a flowchart of the spacecraft image removal process using multi-light angle image fusion according to the present invention.

[0043] Figure 2 This is a schematic diagram of the overall method architecture of the present invention;

[0044] Figure 3 This is a schematic diagram illustrating the process of saving images acquired by the model of the present invention;

[0045] Figure 4 This is a schematic diagram of the weight calculation module structure of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Please see Figure 1-4 This invention provides a technical solution: a method for removing shadows from spacecraft images based on multi-illumination angle image fusion.

[0048] Data collection: Using 3D animation rendering and production software to create composite images of spacecraft;

[0049] Spacecraft Model Acquisition: The dataset uses spacecraft models, which are collected from the Internet. The collected models include different document formats. First, all 3D document formats must be converted to FBX format.

[0050] Rendering process: Unreal Engine generates simulated images that resemble the real scene;

[0051] Methodology Architecture: This paper presents a methodology for reconstructing a shadow-free image from two shadowed spacecraft images through multi-light angle image fusion, and provides an analysis of the mission problem and an explanation of the overall methodology framework.

[0052] Shadow detection: Based on the characteristic that the pixels in the shadow area have low brightness, the input image {I1,I2} is converted into a grayscale image, and the pixel values ​​are compared with the threshold value one by one.

[0053] Image feature extraction and reconstruction: After degrading the image quality, reconstruction is performed to improve the robustness of the encoder and decoder in the face of image noise;

[0054] Weight calculation: The loss function is designed by combining supervised learning and self-supervised learning.

[0055] Experimental Results and Analysis: Encoder-decoder networks were pre-trained on a large natural image dataset.

[0056] Data collection includes spacecraft model acquisition and rendering processes;

[0057] The process of acquiring and saving images of a spacecraft model involves the following steps: assembling a 3D spacecraft model, converting the 3D model format, initial settings for the rendering software, adjusting the lighting angle, rendering the 3D model, and saving the spacecraft image. To simulate the lighting environment in space, after importing the spacecraft 3D model into Unreal Engine 4, atmospheric fog, sky sphere, skylight, and spherical reflection capture were disabled. The light source was set to directional light. By adjusting the angle of the directional light, we can obtain spacecraft images under multiple lighting angles.

[0058] During the rendering process, after importing the 3D model of the spacecraft into Unreal Engine, atmospheric fog, sky sphere, skylight, and spherical reflection capture are turned off. The light source is set to directional light, and the angle of the directional light is adjusted to obtain spacecraft images under multiple illumination angles. The dataset is named "Multi-Illumination Angle Spacecraft Image Set".

[0059] The dataset contains 50 training scenarios and 17 validation scenarios. Each scenario consists of two sequences of spacecraft images from different lighting angles, ground-based images of the spacecraft under uniform lighting conditions, and semantic segmentation masks for labeled spacecraft components.

[0060] The overall method framework requires two spacecraft images under different lighting angles as input, and assumes that these two images have been configured to the same camera viewpoint using an image registration method, where I1∈R. H×W×3 and I2∈R H ×W×3 This represents two images of a spacecraft under shadow at different lighting angles, I f ∈R H×W×3 To obtain a fused spacecraft image free of shadow defects, the spacecraft image fusion task can be described as follows:

[0061] I f =φ(I1,I2) (1)

[0062] In the formula, φ(·) represents the mapping method for reconstructing paired shadowed spacecraft images {I1,I2} into a shadow-free spacecraft image. Our image fusion method architecture's shadow detection module extracts spacecraft image shadows (excluding those against the cosmic background) by comparing the similarities and differences in information between spacecraft images at different illumination angles. The extracted spacecraft image shadow region mask can be represented as {S1,S2}. Therefore, by performing pixel-level multiplication between the input image and the shadow mask as shown in Equation (2), we can obtain the complementary information image {C1,C2} for each input image.

[0063]

[0064] The input shadowed spacecraft image and the compensated image are converted into the YCbCr color space. The luminance channel Y is fused by the proposed network, and the chrominance (Cb and Cr) channels are weighted and fused. The fused luminance and chrominance channels are reconstructed into an RGB image because the image structural details and compensation information are mainly found in the luminance channel.

[0065] Unlike shadow detection in general natural images, shadow areas in spacecraft images are often difficult to distinguish from the dark background of the universe. Therefore, the shadow detection module not only needs to detect the shadow areas in the spacecraft image, but also needs to separate them from the background. The core idea of ​​the module is to compare the differences between two input spacecraft images with different lighting angles.

[0066] First, based on the characteristic of low pixel brightness in shadow areas, the input image is converted to grayscale. Each pixel value is compared to a threshold. For shadow detection, the threshold is set to 50 based on the test results. Pixels with values ​​less than the threshold are considered shadow areas and assigned a value of 1, while the rest are assigned 0. The resulting initial shadow mask is {S1, S2}. In the above steps, the algorithm compares two images. While the background is consistent, the shadow areas are inconsistent. The two initial shadow masks are compared pixel by pixel. Areas with similar brightness are considered the background and assigned a value of 0, while the rest remain unchanged, resulting in the shadow mask after removing the background. Finally, according to formula (3), the complementary information image corresponding to each input image is obtained. In the process of shadow detection, the input image {I1,I2} is converted into a grayscale image, and the pixel values ​​are compared with the threshold (50) one by one. Pixel areas smaller than the threshold are considered as shadow areas, and the shadow areas are assigned a value of 1, while the rest are 0. The initial shadow mask is obtained as {S1,S2}. The differences between the two initial shadow masks are compared pixel by pixel. The same areas are considered as the cosmic background and are assigned a value of 0, while the rest remain unchanged, thus obtaining the shadow mask after removing the background. Input image and shadow mask pair Perform pixel-by-pixel multiplication to obtain the complementary information image corresponding to the input image.

[0067] Image feature extraction and reconstruction: The Y-channel images of the input image and the complementary image are fed into the encoder network to extract features. The output of feature extraction is a feature map. Simultaneously, the input image and its corresponding complementary image are jointly input into the third module, the weight calculation module.

[0068] The weight calculation process can be expressed as:

[0069] W i =H(I i C i ),i∈{1,2} (3)

[0070] H(·) represents the weight calculation function, which is fitted using a three-layer neural network, W i Let represent the fusion weights used to fuse the feature maps of the i-th image and its complementary image. The fusion process can be represented by equation (4). After fusing the two input images with their complementary images respectively, the feature maps are obtained. Figure 1 and characteristics Figure 2 Element-level addition and fusion are performed to obtain the final feature map F. fuse It is reconstructed into an image by the decoder.

[0071]

[0072] First, based on the characteristic of low pixel brightness in shadow areas, the input image {I1, I2} is converted to a grayscale image. Each pixel value is compared to a threshold. Based on the test results, the threshold is set to 50. Pixels with values ​​less than the threshold are considered shadow areas and assigned a value of 1, while the rest are assigned 0, resulting in the initial shadow mask {S1, S2}. Next, the two initial shadow masks are compared pixel by pixel. Areas with similar values ​​are considered the cosmic background and assigned a value of 0, while the rest remain unchanged, resulting in the shadow mask after removing the background. Finally, according to formula (3), the complementary information image corresponding to each input image is obtained.

[0073] In the weight calculation, a loss function is designed by combining supervised learning and self-supervised learning. The input image pairs {I1,C1} and {I2,C2} can obtain fusion weights W1 and W2 through the weight calculation network. After the fusion process of the above equation (7), the fusion feature F can be obtained. fuse_1 F fuse_2 and F fuse These three fused features are all reconstructed by the decoder to obtain the output image I. out1 I out2 and I outFirst, the image similarity loss between the two images is calculated. This is a self-supervised loss that can be used to constrain the brightness distribution of shadow and non-shadow regions to be consistent. Then, the image similarity loss between the two images and the ground truth image with good lighting is calculated to constrain the global image quality after weight fusion. I is then calculated. out True value of the same well-lit image I G Image similarity loss between them;

[0074] Structural similarity loss measures the structural similarity between the input image and the reconstructed image, helping the network to better preserve image structural information. Its calculation formula is as follows:

[0075] L ssim =1-SSIM(I in ,I out (5)

[0076] The network loss function for weighted calculation can be expressed as:

[0077] Loss = L ssim (I out1 ,I out2 )+L ssim (I out ,I G (6)

[0078] The encoder-decoder network was pre-trained in the large natural image dataset MS-COCO2014

[19] and followed the implementation details such as the loss function and training steps in TransMEF. Then, the encoder-decoder network parameters in the MIAS dataset were fine-tuned to better learn spacecraft image features. The weighted sub-network blocks were trained on the MIAS dataset with batch sizes of 16 and 700 epochs. The size of all images was adjusted to 256x256. During training, the ADAM optimizer and cosine annealing learning rate adjustment strategy were used. The learning rate was set to 1e-4 and the weight decay was 0.0005.

[0079] During the evaluation process, the shadow removal results were evaluated using the validation scenarios in the MIAS dataset. We compared our method with eight competitive traditional image fusion methods and deep learning-based image fusion methods. The traditional methods included DME

[20] , DSIFT

[21] , FMMEF

[22] , MEFDISFT

[23] , and the deep learning-based methods included Deepfuse[9], IFCNN

[10] , TransMEF

[11] , and U2Fusion

[12] . Deepfuse, U2Fusion, and TransMEF are all unsupervised learning methods, while IFCNN is based on supervised learning methods.

[0080] To verify that the proposed MIAF task can improve the accuracy of spacecraft component recognition, the performance of the fusion results on the spacecraft semantic segmentation task was studied.

[0081] The deeplabv3+

[24] semantic segmentation network was trained on the spacecraft image dataset, while the images in the MIAF dataset were only used for the verification stage. The quantitative comparison results are shown in Table 1. We used the overall pixel precision (PA), average pixel precision (MPA) and average intersection-union ratio (MIoU) to evaluate the segmentation performance of different methods. The proposed image fusion architecture in the table achieved the best value under all evaluation metrics. The comparison of the segmentation results also shows that, since these image fusion methods destroy the original image structure information, inappropriate image fusion methods will reduce the segmentation accuracy of spacecraft images.

[0082] Table 1

[0083]

[0084] To verify the effectiveness of the specific design weighted sub-blocks, two ablation studies were conducted. The first ablation study directly fused the aligned input images without calculating the compensation image. The second ablation study used only the ground truth image as supervision and did not perform self-supervision in the loss function. Two examples of fusion results in the ablation study show that direct fusion results cannot eliminate shadows in spacecraft images because it does not know the location of the shadow areas. Self-supervised loss can reduce the noise of the fused image and soften the edges of the shadow areas.

[0085] To quantify the effectiveness of the ablation study, semantic segmentation verification was performed. The comparison results are shown in Table 2, verifying that the specifically designed weighted sub-blocks not only achieved better visual performance but also improved the accuracy of spacecraft semantic segmentation.

[0086] Table 2

[0087]

[0088] This method creates publicly available spacecraft image datasets under multiple lighting angles and proposes a novel image fusion framework for shadow removal. Unlike handcrafted image fusion strategies, it designs a weighted sub-network block to learn the optimal feature fusion strategy to reduce the differences in lighting intensity distribution in the input spacecraft images, thereby removing image shadows. It combines supervised loss and self-supervised loss functions to constrain the consistency of brightness distribution in shadow and non-shadow areas.

[0089] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for removing shadows from spacecraft images based on multi-illumination angle image fusion, characterized in that: The method includes: Data collection: Use 3D animation rendering and production software to create synthetic spacecraft images; use spacecraft models in the dataset, which are collected from the Internet. The collected models include different document formats, and all 3D document formats must first be converted to FBX format. Rendering process: Generating simulated images that resemble real-world scenes using Unreal Engine; Shadow detection: Based on the characteristic of low pixel brightness in shadow areas, the input image is... The image is converted to grayscale. Each pixel value is compared to a threshold. Based on the test results, the threshold is set to 50. Pixels with values ​​less than the threshold are considered shadow areas and assigned a value of 1, while the rest are assigned 0. The resulting initial shadow mask is {S1, S2}. The two initial shadow masks are then compared pixel by pixel. Identical areas are considered the background and assigned a value of 0, while the rest remain unchanged, resulting in the shadow mask after removing the background. ; Complementary image generation: Pairing the input image with a shadow mask , Perform pixel-by-pixel multiplication to obtain the complementary information image corresponding to the input image. ; Color space conversion: The input shadowed spacecraft image and information complementary image are converted into the YCbCr color space. The luminance channel Y is fused by the encoder-decoder network, and the chrominance channels Cb and Cr are weighted and fused. The fused luminance and chrominance channels are reconstructed into an RGB image. Image feature extraction and reconstruction: The Y-channel images of the input image and the complementary image are input into the encoder network to extract features. The output of feature extraction is a feature map. Simultaneously, the input image and its corresponding complementary image are jointly input into the weight calculation module; Weight calculation: The weight calculation process can be represented as follows: (1); This represents the weight calculation function, which is fitted using a three-layer neural network. Let represent the fusion weights used to fuse the feature maps of the i-th image and its complementary image. The fusion process can be represented by equation (2). After fusing the two input images with their complementary images respectively, feature map 1 and feature map 2 are obtained. These are then fused element-wise to obtain the final feature map. The spacecraft image was reconstructed by the decoder to be without shadows; （2） Network training: The encoder-decoder network is pre-trained on a large natural image dataset.

2. The method for removing shadows from spacecraft images based on multi-illumination angle image fusion according to claim 1, characterized in that: The data collection includes spacecraft model acquisition; the process of acquiring and saving the spacecraft model image is as follows: spacecraft 3D model acquisition, 3D model format conversion, initial settings of rendering software, lighting angle adjustment, rendering of 3D model, and saving of spacecraft image.

3. The method for removing shadows from spacecraft images based on multi-illumination angle image fusion according to claim 1, characterized in that: During the rendering process, after importing the 3D model of the spacecraft into Unreal Engine, atmospheric fog, sky sphere, skylight, and spherical reflection capture are turned off, and the light source is set to directional light. Adjusting the angle of the directional light, spacecraft images under multiple illumination angles are obtained, and the dataset is named the Multi-Illumination Angle Spacecraft Image Set. The dataset contains 50 training scenarios and 17 validation scenarios. Each scenario consists of two sequences of spacecraft images at different illumination angles, ground-based spacecraft images under uniform illumination conditions, and semantic segmentation masks for labeled spacecraft components.

4. The method for removing shadows from spacecraft images based on multi-illumination angle image fusion according to claim 1, characterized in that: The weight calculation incorporates a loss function design that combines supervised and self-supervised learning methods, with the input image pair... and The fused weights can be obtained through a weight calculation network. and After the fusion process in equation (2) above is performed, the fusion characteristics can be obtained. , and These three fused features are all reconstructed by the decoder to obtain the output image. , and First calculate and The image similarity loss between the two regions is a self-supervised loss that can be used to constrain the brightness distribution of shadowed and unshadowed regions to be consistent; then the calculation is... True value of images with good lighting Image similarity loss is used to constrain the global image quality after weight fusion; Structural similarity loss measures the structural similarity between the input image and the reconstructed image, helping the network to better preserve image structural information. Its calculation formula is as follows: （3） in Indicates the input image. This represents the reconstructed output image; The network loss function for weighted calculation can be expressed as: （4）。

Images