Image deblurring method, system and device for space target

Abstract

The invention discloses a space target-oriented image deblurring method, system and device, and the method comprises the steps: constructing a deblurring model which takes UNet as a backbone network, and integrating a perception adapter at a decoder part of the deblurring model; an initial blurred image is input into the model. Firstly, an encoder is used for carrying out feature extraction and encoding on an image, and a decoder is combined with encoding features and enhancement processing of a perception adapter to gradually reconstruct and output a final clear image. Primarily enhancing a focusing feature of a branch I of the sensing adapter and extracting key information; the branch II takes the output of the branch I as the input to supplement details; and the final output of the sensing adapter adopts a double-branch adaptive weighted fusion strategy. The system comprises an image acquisition unit, an encoder module and a decoder module. According to the invention, both the deblurring precision and the calculation complexity are considered. The method can be widely applied to the field of image processing.

Description

A method, system, and apparatus for deblurring images of space targets. Technical Field

[0001] This invention relates to the field of image processing, and in particular to an image deblurring method, system, and apparatus for spatial targets. Background Technology

[0002] With the rapid advancement of space technology, the number of satellites in orbit has surged. Simultaneously, the cumulative amount of failed satellites and debris is also increasing exponentially, creating a trend of "space fragmentation." This poses significant risks to subsequent launches, on-orbit operations, and collision avoidance. Visual perception is a core component for achieving high-precision docking, debris capture, and satellite status monitoring. As countries advance on-orbit services, space operations rely on high-precision visual understanding of critical components of space targets. However, actual space target imaging is affected by factors such as target motion, camera vibration, and defocusing, easily leading to image blurring and loss of detail, which in turn causes target positioning errors, category confusion, and even collision risks. Furthermore, existing image deblurring methods are overly complex; small kernel stacking or complex operators introduce high latency and complexity challenges, making them difficult to deploy on mobile devices and therefore unsuitable for space targets. Summary of the Invention

[0003] In view of this, to address the technical problem of balancing deblurring accuracy and computational complexity in existing space target imaging methods, this invention proposes an image deblurring method for space targets. This method includes the following steps: First, the blurred image to be processed is acquired as input. Then, a deblurring model with UNet as the backbone network is constructed, and a perceptual adapter module is integrated into its decoder. During processing, the initial blurred image is input into the model. The model first uses the encoder to extract and encode features from the image, obtaining a deep feature representation. Next, the decoder combines the encoded features with the enhancement processing of the perceptual adapter to gradually reconstruct and output the final clear image.

[0004] The perceptual adapter performs initial enhancement of branch-focused features and extraction of key information, including normalization and preprocessing with 1×1 convolution, precise localization with 3×3 depth-separable convolution, and depth enhancement with SCMA layer.

[0005] The second branch of the perception adapter takes the output of the first branch as input and supplements the details, including a secondary enhancement that inherits the SCMA Layer.

[0006] The final output of the perception adapter employs a dual-branch adaptive weighted fusion strategy.

[0007] The present invention also proposes an image deblurring system for space targets, which includes an image acquisition unit, an encoder module and a decoder module.

[0008] The present invention also proposes an image deblurring device for space targets, comprising: at least one processor; at least one memory for storing at least one program; when the at least one program is executed by the at least one processor, the at least one processor implements the image deblurring method for space targets as described above.

[0009] Based on the above scheme, this invention provides an image deblurring method, system, and device for spatial targets. It designs an effective perception adapter (Per-Adapter) that addresses the contradiction between computational complexity and detail reconstruction accuracy in existing image deblurring methods through coordinated optimization of lightweight architecture design and scene-based adaptive feature modulation. This method is adapted to the characteristics of strong local correlation and high detail requirements of spatial target images. The core of this method is a mobile-friendly large-kernel depthwise separable convolution. It innovatively employs large-kernel depthwise separable convolution, channel splitting, and staggered projection to construct basic components, ensuring efficient feature mixing while overcoming the bottlenecks of limited receptive field and insufficient detail capture in traditional lightweight models. Furthermore, through dual-branch iterative enhancement and adaptive weighted fusion architecture, it achieves a step-by-step feature optimization of "coarse-fine tuning," improving the clarity of spatial target images while reducing model computational complexity and FLOPs, thus balancing lightweight deployment with high accuracy requirements. Attached Figure Description

[0010] Figure 1 is a flowchart of the steps of an image deblurring method for spatial targets according to the present invention; Figure 2 is a schematic diagram of the data flow of the deblurring model in a specific embodiment of the present invention; Figure 3 is a structural block diagram of the perception adapter module in a specific embodiment of the present invention; Figure 4 is a schematic diagram of the deblurring visualization result of the spatial target image in a specific embodiment of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] It should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0013] It should be understood that the terms "system," "apparatus," "unit," and / or "module" used in this application are a method of distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0014] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "a," and / or "the" are not specifically singular and may include the plural. Generally, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements. An element defined by the phrase "comprising an..." does not exclude the presence of other identical elements in the process, method, product, or apparatus that includes the element.

[0015] In the description of the embodiments of this application, "a plurality of" refers to two or more. The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0016] Furthermore, flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, the steps can be processed in reverse order or simultaneously. Additionally, other operations can be added to these processes, or one or more steps can be removed from them.

[0017] Referring to Figure 1, a flowchart illustrating an optional example of the image deblurring method for spatial targets proposed in this invention is shown. This method can be applied to computer devices. The image deblurring method proposed in this embodiment may include, but is not limited to, the following steps: Step S1, acquiring an initial image; Step S2, constructing a deblurring model by introducing a perceptual adapter in the decoder based on the UNet architecture; Step S3, inputting the initial image into the deblurring model; Step S4, encoding the initial image through an encoder to obtain encoded data; Step S5, decoding the encoded data through a decoder to output a clear image.

[0018] The model employs a NAFNet module as the encoder and combines it with a Per-Adapter module as the decoder. The specific data flow is shown in Figure 2. First, the blurred image undergoes 3×3 convolution processing before being input to the encoder. The image is then downsampled through multiple NAFNet blocks in the encoder, and data of the corresponding size is output (the size of this batch of data includes...). , , , , Then, it jumps to the corresponding perceptual adapter of the decoder and the multi-layer NAFNet Block, where, The data of this size is further processed by an NAFNet block before being concatenated with the NAFNet block of the decoder. The decoder output is then processed by a 3×3 convolution and added pixel-by-pixel to the blurred image to obtain... A clear image of the size.

[0019] In some feasible embodiments, the Per-Adapter module has a dual-branch structure, specifically including: Branch 1 of the Per-Adapter module focuses on preliminary feature enhancement and key information extraction. Addressing the pain point of existing lightweight models either reducing computational cost at the expense of detail or preserving detail while increasing computation, three layers of optimization are designed in Branch 1: preprocessing of normalization and 1×1 convolution: first, normalization is used to stabilize the data distribution, and then 1×1 convolution is used to achieve efficient integration of cross-channel information, solving the problems of inefficient channel interaction and data distribution fluctuation affecting feature quality in traditional models.

[0020] Precise localization with 3×3 depth separable convolution: Compared with traditional standard convolution, this operation focuses on capturing spatial details such as edges, textures, and local contours of spatial target images with single-channel independent convolution. While maintaining the ability to extract spatial features, it reduces the number of parameters and computational complexity, and solves the core contradiction between spatial detail extraction and lightweighting.

[0021] Deep Enhancement of SCMA Layer: Overcoming the limitations of computational redundancy caused by full-channel processing in existing attention mechanisms, a lightweight strategy of single-path enhancement and single-path retention is designed. After the features are evenly divided along the channels, only one path learns complex channel correlations through a closed-loop SBCR module consisting of two 1×1 convolutions and SiLU activations, while the other path directly retains the original information, reducing computational overhead by 50%. Subsequently, a 7×7 large-kernel depthwise separable convolution is used to expand the receptive field, specifically adapting to strong local correlations in spatial target images. This is more effective than the traditional 3×3 kernel in capturing global local structural correlations. Finally, through the SCA channel attention mechanism, global average pooling is used to generate channel weights, which enhances high-frequency key channels such as edges and textures and suppresses redundant channels. This solves the problems of weak feature modulation targeting and key information being masked by redundant signals in existing methods.

[0022] In the image deblurring method, the Per-Adapter module's second branch takes the output of the first branch as input. To address the issues of computational waste and insufficient fine-tuning focus caused by structural repetition in existing dual-branch models, a division-of-labor optimization design is adopted: omitting the 3×3 depthwise separable convolution for precise load reduction: Based on the division-of-labor logic where the first branch has already completed the initial extraction of spatial details, the second branch omits the 3×3 depthwise separable convolution layer, avoiding redundant computation caused by excessive extraction of spatial features. At the same time, computational resources are concentrated on channel-level feature fine-tuning, solving the problems of structural redundancy and low fine-tuning focus in traditional dual-branch models.

[0023] Inheriting the secondary enhancement of SCMA Layer: the innovative logic of fully reusing single-path enhancement, single-path preservation, large kernel receptive field expansion, and SCA attention modulation is used to perform secondary optimization on the features of branch one. On the one hand, it further enhances the subtle and key features such as spacecraft texture and component edges, and on the other hand, it suppresses the unfiltered redundant information of branch one, so that the feature discriminativeness and accuracy are improved compared with the single-branch model.

[0024] The per-Adapter module design in the image deblurring method breaks through the existing simple repetitive enhancement mode of two branches. By dividing the work of spatial extraction in branch one and channel fine-tuning in branch two, the accuracy of detail restoration is significantly improved, and the problem of balancing iterative enhancement and computational efficiency is solved.

[0025] Specifically, the overall structure of the sensing adapter is shown in Figure 3.

[0026] In some feasible embodiments, the final output of the Per-Adapter module adopts a dual-branch adaptive weighted fusion strategy, which breaks through the limitations of existing methods such as fixed weight fusion or simple splicing. In this strategy, two sets of independent learnable parameters dynamically adjust the branch contribution: the weight parameters are adaptively optimized according to the texture complexity and noise level of the space target image. For example, in the spacecraft texture-dense region, the fine-tuned feature weight of branch two is automatically increased, while in the simple region, the coarse-tuned feature weight of branch one is dominant. This solves the problems of poor adaptability of fixed fusion strategies and insufficient feature complementarity in different scenarios.

[0027] The comprehensive advantage of the fused features: Through the synergy of branch one to preserve the structure and branch two to supplement the details, the final output features have both structural integrity and detail discriminability. Compared with the single feature output of existing methods, the support capability for high-dimensional detail restoration such as spacecraft texture completion and component edge sharpening is significantly enhanced.

[0028] Based on the overall process of the above method, this invention also provides relevant experimental examples. In the publicly available space target dataset "A Spacecraft Dataset for Detection, Segmentation and PartsRecognition," which includes 3117 pairs of images and masks at a resolution of 1280×720, and encompasses three categories: main body, solar panels, and antennas, target motion blur, camera shake blur, and defocus blur are added to this dataset to create a space target image deblurring dataset. The designed method was trained and validated and compared with current state-of-the-art image deblurring algorithms. After training on an NVIDIA RTX 4090 single card, performance was compared on the validation set, and the results are shown in Table 1.

[0029] Table 1. Performance metrics of each image deblurring model on the test set. As can be seen from the table above, the method of the present invention demonstrates a significant advantage in balancing the deblurring quality and computational efficiency of spatial target images: the image restoration quality of the method of the present invention is top-notch, with a PSNR of 33.90 and an SSIM of 0.9552. These two core quality indicators rank first among all models in the table above. It has reached the excellent level of the current test set in terms of pixel-level accuracy and structural similarity restoration, and can restore the real details of spatial target images to the greatest extent.

[0030] The method of this invention has high computational efficiency, with MACs of 65.19G, which is within a reasonable range and far lower than models such as MPRNet and Stripformer. The computational load of the method of this invention is close to that of NAFNet, but it achieves both PSNR and SSIM, avoiding excessive computational overhead caused by pursuing ultimate quality, and maintaining high computational efficiency while achieving top-notch recovery quality.

[0031] To clarify the individual contribution and synergistic effect of SCMA Layer image deblurring performance in branch 1 and branch 2, this study analyzed the results through ablation experiments, as shown in Table 2.

[0032] Table 2. Experimental results of SCMA Layer ablation Baseline performance without branches: When the SCMA Layer is not enabled in branches one and two, the model PSNR is 33.63 and SSIM is 0.9531. This is the performance baseline.

[0033] The individual contribution of branch one: When the SCMA layer is enabled only in branch one, the PSNR is improved to 33.66 and the SSIM is improved to 0.9534, indicating that branch one can slightly improve the pixel accuracy and structural similarity of spatial target image deblurring.

[0034] The separate contribution of branch two: When the SCMA layer is enabled only in branch two, the PSNR is further improved to 33.76 and the SSIM is improved to 0.9539, indicating that enabling the SCMA layer in branch two has a more significant effect on the image deblurring quality and a stronger effect on the restoration of structural details (such as spacecraft textures and component edges).

[0035] Dual-branch synergistic effect: When SCMA Layer is enabled in both Branch 1 and Branch 2 simultaneously, PSNR jumps to 33.90 and SSIM reaches 0.9552, achieving a significant performance breakthrough. This indicates that there is a complementary synergistic mechanism between Branch 1 and Branch 2, and enabling both simultaneously can greatly enhance the SCMA Layer's ability to reconstruct details of spatial target images.

[0036] In summary, enabling SCMA Layer in both Branch 1 and Branch 2 can effectively improve image deblurring performance. When the two are used together, they can produce even better results. This is the core design of SCMA Layer for achieving high-performance image deblurring.

[0037] The deblurring visualization results of some spatial target images in the experiment are shown in Figure 4.

[0038] In summary, this invention designs a Per-Adapter, which addresses the contradiction between computational complexity and detail reconstruction accuracy in existing image deblurring methods through coordinated optimization of lightweight architecture design and scene-based adaptive feature modulation. Addressing the high latency and complexity challenges caused by small kernel stacking or complex operators, this invention uses mobile-friendly large kernel depthwise separable convolution as its core, innovatively employing large kernel depthwise separable convolution, channel splitting, and staggered projection to construct basic components. This ensures efficient feature mixing while overcoming the bottlenecks of limited receptive field and insufficient detail capture in traditional lightweight models. It achieves refined detail mining and redundancy suppression through iterative enhancement via a dual-branch approach, overcoming the limitations of fixed-weight fusion or simple splicing in existing methods. This enables step-by-step feature synergy optimization of coarse and fine tuning. Ultimately, compared to typical image deblurring methods, it maintains superior image clarity while reducing model computational complexity and FLOPs, balancing lightweight deployment with high-precision image deblurring requirements.

[0039] Applying the above method, an image deblurring system for spatial targets includes: an image acquisition unit for performing step S1; a model building unit for performing step S2; a data input unit for performing step S3; an encoder module for performing step S4; and a decoder module for performing step S5.

[0040] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0041] An image deblurring apparatus for space targets includes: at least one processor; at least one memory for storing at least one program; when the at least one program is executed by the at least one processor, the at least one processor implements the image deblurring method for space targets as described above.

[0042] The content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0043] A storage medium storing processor-executable instructions, which, when executed by a processor, are used to implement an image deblurring method for spatial targets as described above.

[0044] The content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0045] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for deblurring images of spatial targets, characterized in that, Includes the following steps: Get the initial image; Based on the UNet architecture, a perceptual adapter is introduced into the decoder to construct a deblurring model; the initial image is input into the deblurring model; the initial image is encoded by an encoder to obtain encoded data; the encoded data is decoded by a decoder to output a clear image.

2. The image deblurring method for spatial targets according to claim 1, characterized in that, The NAFNet module is used as the encoder, and the NAFNet module and the perception adapter are used as the decoder.

3. The image deblurring method for spatial targets according to claim 2, characterized in that, The perception adapter includes a first branch and a second branch, wherein: in the first branch, normalization and 1×1 convolution preprocessing, precise localization of 3×3 depth separable convolution and depth enhancement processing of SCMA layer are performed; in the second branch, normalization and 1×1 convolution preprocessing and SCMA layer depth enhancement processing are performed.

4. The image deblurring method for spatial targets according to claim 2, characterized in that, The output of the perception adapter adopts a dual-branch adaptive weighted fusion strategy.

5. The image deblurring method for spatial targets according to claim 1, characterized in that, Also includes: Before the encoder, it undergoes a 3×3 convolution process; after the decoder, it undergoes a 3×3 convolution process.

6. The image deblurring method for spatial targets according to claim 1, characterized in that, Also includes: A training set is constructed and the deblurring model is trained.

7. An image deblurring system for spatial targets, characterized in that, include: The image acquisition unit is used to acquire the initial image; The encoder module is used to encode the initial image; The decoder module introduces a perceptual adapter to decode the encoded data.

8. An image deblurring device for spatial targets, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements an image deblurring method for space targets as described in any one of claims 1-6.

Images