Methods, apparatus, equipment, and storage media for recovering remote sensing images

Abstract

This application relates to the field of image processing technology and discloses a method for restoring remote sensing images, comprising: extracting features from a degraded image to obtain a feature map of the degraded image, and identifying the feature map of the degraded image to determine the key information regions of the degraded image. The degraded image is then input into an image restoration network model, enabling the image restoration network model to identify the key information regions of the degraded image and output a restored image. This method can improve the efficiency of image restoration. This application also discloses an apparatus, device, and storage medium for restoring remote sensing images.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method, apparatus, device, and storage medium for recovering remote sensing images. Background Technology

[0002] Image degradation refers to the phenomenon of image quality deterioration during image acquisition, transmission, or processing due to defects in the imaging system, external interference, or environmental factors. It is commonly seen in scenarios such as underwater scattering, low light noise, atmospheric haze, and motion blur.

[0003] In related technologies, a uniform feature modeling strategy is typically employed during feature extraction when restoring degraded images, implicitly assuming that all regions of the image contribute equally to the restoration process. However, this assumption introduces a large amount of redundant computation into the degraded image, especially the computation related to the background or low-information regions, thereby reducing the efficiency of image restoration. Summary of the Invention

[0004] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0005] This application provides a method, apparatus, device, and storage medium for recovering remote sensing images, which can improve the efficiency of image recovery.

[0006] In a first aspect, embodiments of this application provide a method for recovering remotely sensed images, including: Feature extraction is performed on the degraded image to obtain its feature map; Identify feature maps of degraded images to determine key information regions within the degraded images; The degraded image is input into the image restoration network model, which then identifies key information regions of the degraded image and outputs the restored image.

[0007] Optionally, the feature map of the degraded image is identified to determine the key information region of the degraded image, including: for each feature vector in the feature map, predicting the probability that the feature vector contains key information to obtain the classification result of each spatial location in the feature map; based on the classification result of each spatial location in the feature map, semantically reordering the feature sequence of the feature map to obtain the key information region of the degraded image.

[0008] Optionally, after determining the key information regions of the degraded image, the method further includes: assigning corresponding state transition parameters to different key information regions in the degraded image.

[0009] Optionally, the image restoration network model identifies key information regions of the degraded image and outputs the restored image, including: for each key information region of the degraded image, multi-scale encoding of the key information regions to obtain multi-scale input features; weighted fusion of the multi-scale input features to obtain fused features; multi-scale decoding of the fused features to obtain image prediction results at different scales; and applying pixel-level supervision to the image prediction results at different scales to realize the restoration of the degraded image and obtain the restored image.

[0010] Secondly, embodiments of this application provide an apparatus for recovering remotely sensed images, comprising: The image processing module is used to extract features from the degraded image to obtain the feature map of the degraded image; The region recognition module is used to identify the feature map of the degraded image in order to determine the key information region of the degraded image; The image restoration module is used to input the degraded image into the image restoration network model, so that the image restoration network model can identify the key information regions of the degraded image and output the restored image.

[0011] Optionally, the region recognition module is specifically used to: predict the probability that each feature vector in the feature map contains key information, and obtain the classification result of each spatial location in the feature map; based on the classification result of each spatial location in the feature map, perform semantic reordering on the feature sequence of the feature map to obtain the key information region of the degraded image.

[0012] Optionally, the region recognition module is also used to: after determining the key information regions of the degraded image, assign corresponding state transition parameters to different key information regions in the degraded image.

[0013] Optionally, the image restoration module is specifically used for: multi-scale encoding of each key information region of the degraded image to obtain multi-scale input features; weighted fusion of the multi-scale input features to obtain fused features; multi-scale decoding of the fused features to obtain image prediction results at different scales; and applying pixel-level supervision to the image prediction results at different scales to realize the restoration of the degraded image and obtain the restored image.

[0014] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory storing program instructions, wherein the processor is configured to execute the method for recovering remotely sensed images in the first aspect when running the program instructions.

[0015] Fourthly, embodiments of this application provide a storage medium storing program instructions, wherein the program instructions, when executed on a computer, perform the method for recovering remote sensing images as described in the first aspect.

[0016] The method, apparatus, device, and storage medium for restoring remote sensing images provided in this application can achieve the following technical effects: In this embodiment, feature regions crucial to restoration quality can be automatically identified from the input degraded image. Through targeted modeling tailored to different degradation types, selected regions are emphasized, while redundant optimization of background and low-impact areas is effectively suppressed, thereby improving the efficiency of image restoration.

[0017] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0018] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are considered similar elements. The drawings do not constitute a limitation of scale, and wherein: Figure 1 This is a schematic diagram illustrating the features of different methods for restoring degraded images provided in the embodiments of this application; Figure 2 This is a schematic diagram of the FocusIR framework to which the method for restoring remote sensing images provided in the embodiments of this application applies; Figure 3 This is a flowchart of a method for recovering remotely sensed images provided in an embodiment of this application; Figure 4 These are graphs showing the quantitative analysis results of different datasets in the experiments provided in this application embodiment; Figure 5 This is a graph showing the results of an ablation experiment provided in an embodiment of this application; Figure 6 This is a schematic diagram of an apparatus for recovering remotely sensed images provided in an embodiment of this application; Figure 7 This is a schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] In real-world environments, object imaging is often affected by a variety of degradation factors, such as color distortion caused by light absorption and scattering, and structural information damage caused by noise. These degradation phenomena severely impair critical structures and fine details in images, thereby affecting the performance of subsequent visual tasks, including object detection, scene understanding, and visual perception.

[0020] For degraded images, related technologies offer methods for image restoration. For example, Transformer-based methods enhance long-range dependency modeling through a global self-attention mechanism. Mamba-based methods achieve global context modeling with linear computational complexity. Diffusion-based methods improve perceptual quality by representing image restoration as a multi-step iterative denoising process.

[0021] However, while related techniques for restoring degraded images employ different modeling paradigms, they share a common characteristic in the feature extraction stage: they typically utilize a dense and uniform modeling strategy, implicitly assuming that all image regions have equal importance. In real-world degraded images, this assumption often contradicts the highly uneven distribution of information, leading to a significant waste of computational resources on the background or low-impact regions, thus limiting computational efficiency and the ability to recover fine details.

[0022] For example, in combination Figure 1 As shown, this paper presents a feature representation for a different method of restoring degraded images. In degraded images, only a small portion of the regions containing structural and semantic information plays a decisive role in the restoration quality, while large background areas contribute negligibly to the final result. However, related methods for restoring degraded images, whether based on self-attention, state-space modeling, or iterative diffusion, repeatedly and equally model all regions during the optimization process. This uniform modeling strategy not only introduces a large amount of computational redundancy but also weakens the model's ability to focus on informative regions.

[0023] In addition, in related technologies, degraded image restoration methods are mainly divided into two categories: image restoration-based and image enhancement-based.

[0024] Traditional image restoration methods typically describe the degradation process based on explicit prior assumptions or physically based imaging models, and solve the inverse problem through optimization to recover the original sharp image. These methods usually formulate image degradation as additive noise contamination, convolutional blurring, or illumination attenuation, and incorporate prior constraints to mitigate the inherent ill-posedness of the restoration problem. Representative priors include total variational regularization, sparse representation, low-rank constraints, and natural image priors derived from statistical distributions. Although these methods have strong physical interpretability, their performance largely depends on the validity of the assumed priors, and they often fail to achieve satisfactory results under complex or unknown degradation conditions.

[0025] Learning-based image restoration methods are applied to tasks such as image denoising, deblurring, dehazing, and super-resolution, benefiting from hierarchical feature extraction and nonlinear mapping techniques. However, due to the local receptive field characteristics of convolutional operations, convolutional neural network (CNN) models still have limited ability to capture long-range dependencies. The Transformer, by explicitly modeling global contextual relationships through a self-attention mechanism, overcomes this limitation. To reduce the high computational complexity of full attention processing for high-resolution images, existing methods typically employ windowed or striped attention schemes to improve efficiency. The Mamba architecture, based on a state-space model, has been introduced into vision tasks, providing near-linear computational complexity for long sequence modeling and offering a promising alternative for efficient long-range dependency modeling. However, most current Mamba-based image restoration methods require converting two-dimensional images into one-dimensional sequences, which often disrupts the coherence of spatial structure, thus reducing restoration effectiveness.

[0026] To enhance the flexibility and scalability of the model, plug-and-play (PnP) image restoration methods are employed. These methods typically divide the restoration process into two parts: a data consistency term and a prior modeling term. The prior can be implemented using traditional regularization functions or deep learning-based models and integrated into existing optimization frameworks or network architectures in a modular manner. Some methods use pre-trained denoising networks as implicit priors to guide the restoration process iteratively, while others design lightweight, pluggable functional modules that can be seamlessly inserted into different image restoration networks without modifying the overall architecture. By significantly improving versatility and adaptability while maintaining restoration performance, PnP methods are particularly suitable for multi-task and multi-degradation restoration scenarios.

[0027] In view of the above, embodiments of this application provide a method, apparatus, device, and storage medium for restoring remote sensing images. The scheme employs a selective image restoration framework called "FocusIR" (Focus Image Restoration). The core of FocusIR lies in a plug-and-play adaptive information region selection module, which automatically identifies regions crucial to the restoration quality of damaged images before feature extraction. Then, through targeted modeling and enhancement tailored to different degradation types, the selected information regions are emphasized, while redundant optimization of the background and low-impact areas is effectively suppressed, thereby improving the efficiency of image restoration.

[0028] To provide a more detailed understanding of the features and technical content of the embodiments of this application, the implementation of the embodiments of this application will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this application. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0029] The terms "first," "second," etc., used in the specification, claims, and drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. Unless otherwise stated, the term "a plurality of" means two or more.

[0030] In this embodiment, the character " / " indicates that the preceding and following objects are in an "or" relationship. For example, A / B means: A or B. The term "and / or" describes an association relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or, A and B. The term "correspondence" can refer to an association or binding relationship; A corresponding to B means that there is an association or binding relationship between A and B.

[0031] Combination Figure 2 As shown, the FocusIR framework used in the method for restoring remote sensing images provided in this application mainly includes two core components: (1) a plug-and-play Adaptive Informative Region Selection (AIRS) module, which can identify, select and reorganize regions containing key information before feature modeling; and (2) a Multi-scale Image Restoration Network (MIRN) model, which progressively enhances and reconstructs selected features in a hierarchical manner.

[0032] Based on the aforementioned FocusIR framework, this application provides a method for recovering remote sensing images, combined with... Figure 3 As shown, the method includes the following steps: S31, perform feature extraction on the degraded image to obtain the feature map of the degraded image.

[0033] S32, Identify the feature map of the degraded image to determine the key information region of the degraded image.

[0034] S33, input the degraded image into the image restoration network model so that the image restoration network model can identify the key information regions of the degraded image and output the restored image.

[0035] In the method for restoring remote sensing images provided in this application embodiment, feature maps of the degraded image are obtained by feature extraction and identification of the degraded image's feature maps to determine key information regions of the degraded image. Feature regions crucial to restoration quality are automatically identified from the input degraded image. The degraded image is then input into an image restoration network model, enabling the model to identify key information regions of the degraded image and output the restored image. Through targeted modeling tailored to different degradation types, selected regions are emphasized, while redundant optimization of background and low-impact regions is effectively suppressed, thereby improving the efficiency of image restoration.

[0036] The following describes the optional implementation methods in steps S31 to S33.

[0037] Optionally, in step S32 above, identifying the feature map of the degraded image to determine the key information regions of the degraded image includes: for each feature vector in the feature map, predicting the probability that the feature vector contains key information, and obtaining the classification result for each spatial location in the feature map. Based on the classification result for each spatial location in the feature map, semantically reordering the feature sequence of the feature map to obtain the key information regions of the degraded image. After determining the key information regions of the degraded image, assigning corresponding state transition parameters to different key information regions in the degraded image.

[0038] In this implementation, key information regions in degraded images can exhibit stronger structural changes and more significant semantic responses. Therefore, instead of performing equal-weighted modeling for all spatial locations, this implementation estimates the importance of regions from feature embeddings. The core objective of AIRS is to explicitly identify and prioritize regions that play a decisive role in the recovery quality before feature modeling, thereby guiding subsequent networks to focus their representational capabilities on key structures and fine details.

[0039] Given the feature maps F of a degraded image as input, rearrange them into a sequence. The specific formula is as follows: In the above formula, Represents the first in the feature map 1 eigenvector Indicates the height of the feature map. Indicates the width of the feature map. This indicates the total length of the feature sequence.

[0040] AIRS employs a lightweight routing network to predict the information class assignment for each feature vector in the feature map, i.e., each spatial location. For each feature vector... Its score The calculation formula is as follows, and the score is... Used to characterize the probability that each feature vector contains key information: The above formula is then used to obtain the probability distribution via the Softmax function. The details are as follows: To achieve differentiable discrete selection, the Gumbel-Softmax approximation is used to obtain the hard assignment result. ,in, This represents the probability that the i-th spatial location is assigned to one of the K key information regions. AIRS performs semantic reordering on the feature sequence, arranging features with similar information importance into a continuous sequence. The specific formula is as follows: Through the above calculation process, spatial tokens with similar information importance can be aggregated together, which is beneficial for the subsequent image restoration network model to carry out efficient information dissemination in key information areas and reduce interference from background areas.

[0041] A selective scanning mechanism based on the State-Space Model (SSM) is introduced on the reordered feature sequence, with the specific formula as follows: In the above formula, , Indicates the first State transition parameters for class information regions This represents the state value at time t. These are the reordered eigenvalues ​​at time t.

[0042] Different information categories correspond to different degradation characteristics and structural complexities. Assigning region-specific state transition parameters to different regions enables more flexible and adaptive modeling dynamics. This mechanism allows image restoration network models to strengthen the state update process in key information regions while suppressing interference from background regions during state propagation.

[0043] Finally, the modeled features are mapped back to the original spatial structure through an inverse reordering operation to obtain the enhanced feature map. The specific formula is as follows: In the above formula, Y represents the feature sequence after selective scanning. Gamma correction can reduce the brightness of blurred areas without losing content details. However, gamma correction is a global contrast enhancement operation; because it ignores local detail information, it calculates a locally contrast-enhanced version of the initial dehazed image by appropriately enhancing the high-frequency components of the brightness channel.

[0044] Optionally, in step S33 above, the image restoration network model identifies key information regions of the degraded image and outputs the restored image, including: for each key information region of the degraded image, multi-scale encoding is performed on the key information regions to obtain multi-scale input features. The multi-scale input features are weighted and fused to obtain fused features. The fused features are then decoded at multiple scales to obtain image prediction results at different scales. Pixel-level supervision is applied to the image prediction results at different scales to achieve the restoration of the degraded image, resulting in the restored image.

[0045] In this implementation, after determining the key information regions of the degraded image, the enhanced feature map is input into a nested multi-scale network of MIRN for restoration. This network employs an encoder-decoder architecture and achieves progressive enhancement through multi-scale input and multi-scale output. The formula for the encoding process at the s-th scale is as follows: In the above formula, This represents a feature extraction unit at the s-th scale consisting of a series of residual blocks. This represents the input feature at the s-th scale.

[0046] To enhance cross-scale information transmission, this implementation incorporates multiple feature fusion modules into the network, with the calculation formula as follows: In the above formula, This indicates that it is supplementing low-level contextual information. This indicates the relative importance of features at different scales. This represents the activation function.

[0047] Adaptive feature fusion yields fused features, and the specific calculation formula is as follows: In the above formula, Indicates the characteristics after fusion. Indicates the first Weight coefficients for each scale feature.

[0048] Given that the image restoration network model employs a multi-scale output architecture, this implementation applies pixel-level supervision to the prediction results at different scales. For example, the network outputs at the three scales are represented as follows: The corresponding downsampled real image is .

[0049] The formula for the multi-scale pixel domain reconstruction loss is as follows: In the above formula, This represents the network output image at the s-th scale. This represents the true image after downsampling at the s-th scale. It is an L1 norm.

[0050] To further illustrate the beneficial effects of the method for restoring remote sensing images provided in the embodiments of this application, an experimental scheme is provided to verify the method for restoring remote sensing images provided in the embodiments of this application, as follows: Experimental Datasets: Six datasets were used to evaluate the restoration performance of the proposed method, covering tasks such as underwater enhancement, image dehazing, low-light enhancement, and image deblurring. Specifically, two underwater datasets were used. The UCCS dataset consists of three subsets, each containing 100 images with green, blue, and blue-green hues, primarily used to evaluate color correction capabilities in underwater image enhancement. The UIEB dataset contains 890 underwater images collected from different degraded scenes and is widely used to evaluate the overall enhancement performance of different methods. For image dehazing, two datasets were used. The UAV dataset is a partially publicly available hazy dataset designed to evaluate the dehazing performance of various methods. The SIHI-WHU dataset provides robustness evaluation of fog removal methods in different scene environments. In addition, a low-light dataset and a deblurring dataset were included. The LOM dataset was used to evaluate the algorithm's performance in different restoration tasks, while the GoPro dataset contains 2103 images for training and 1111 images for testing; this dataset is widely used in image deblurring evaluation.

[0051] Baseline approach for the experiment: A comprehensive comparison of five representative methods was conducted for each task.

[0052] In underwater image enhancement, the proposed method is compared with FGAN, SMDR, CCL, Water-Net, and PUGAN.

[0053] In terms of image dehazing, the proposed method was compared with CGID, DNMGDT, PMT, SSID and DEA-Net.

[0054] In terms of low-light image enhancement, the proposed method is compared with DiffLL, GSAD, UPT-Flow, PHP and PairLLE.

[0055] In terms of image deblurring, the proposed method is compared with DMPHN, SAPHN, NAFNet, GRL and EVSSM.

[0056] Experimental Details: To ensure fairness in comparing the remote sensing image restoration method provided in this application with other degraded image restoration methods, the results reported by AFAM and all comparison methods were performed on the same device and under the same configuration settings. For baseline methods, their source code was used and the reproduced performance was reported. For deep learning-based methods, the ADAM optimizer with default parameters and an initial learning rate of 0.0001 was used. This experiment was conducted in a PyCharm 2022 environment using Python 3.7 and CUDA 11.6. For traditional methods, this experiment was implemented using MATLAB R2020a.

[0057] Experimental metrics: Different evaluation metrics were used for each restoration task. For underwater image enhancement, two no-reference quality metrics, UIQM and UCIQE, were used. For image dehazing, two no-reference quality metrics, AG and EI, were used. For low-light enhancement and image deblurring, the widely adopted full-reference metrics PSNR and SSIM were used to evaluate performance.

[0058] Quantitative comparison of experiments: combined Figure 4As shown, to evaluate the effectiveness of the proposed method under different restoration types, qualitative and quantitative analyses were performed on four representative degradation scenarios. For underwater image enhancement, the proposed method demonstrated consistent superiority in both qualitative and quantitative evaluations. Compared with representative methods such as Water-Net and CCL, the proposed method produces more natural color correction, higher contrast, and clearer structural details, while effectively avoiding oversaturation and color distortion. Quantitatively, the proposed method achieved the highest UIQM scores on both the UCCS (3.412) and UIEB (3.473) datasets, and the best UCIQE (0.551) score on UCCS. These results indicate that the proposed method is particularly suitable for improving perceptual quality and color fidelity under complex underwater degradation conditions. In image deblurring tasks, the proposed method provides competitive restoration quality and delivers clear visual improvements. Compared with recent methods such as EVSSM, the remote sensing image restoration method provided in the above embodiments produces sharper edges and more coherent textures while suppressing residual blur artifacts. From a quantitative perspective, the proposed method achieves a PSNR of 33.842 dB and an SSIM of 0.9681 on the GoPro dataset, outperforming most competing methods and nearly matching the best-performing one. This demonstrates that the proposed framework effectively preserves detail and structural consistency when removing motion blur. For image dehazing, the proposed method exhibits strong robustness across different scene types. It delivers sharper scene visibility and better contrast than methods such as DNMGDT and CGID, while avoiding over-enhancement in uniform regions. Figure 4 The quantitative results further confirm the following observations: the proposed method achieves the highest AG and EI scores (2.746 and 22.743, respectively) on the UAV dataset and also reaches best performance on the SIHI-WHU dataset, with AG = 7.864 and EI = 81.356. These results highlight the proposed method's ability to effectively handle spatially varied haze and diverse atmospheric conditions. In low-light enhancement tasks, the proposed method consistently improves visibility while maintaining natural lighting effects. Compared to UPT-Flow and other low-light methods, the proposed method recovers sharper object outlines and more uniform brightness without introducing significant noise amplification. Quantitatively, combined with... Figure 4 As shown, the proposed method achieves the highest SSIM value of 0.8113 and a competitive PSNR value of 23.948 dB on the LOM dataset, outperforming most comparable methods. These results demonstrate that the proposed method can effectively enhance low-light images while maintaining structural integrity.

[0059] Ablation experiment: combined Figure 5 As shown, an ablation study was conducted on low-light enhancement and motion deblurring tasks to evaluate the effectiveness and efficiency of AIRS. These two types of degradation differ fundamentally in their information distribution: low-light images typically contain a large number of low-information or noise-dominated regions and sparse useful structures, while motion blur usually affects the entire image and requires globally consistent restoration. This comparison allows for a comprehensive analysis of AIRS performance under different regional information features, including restoration quality and computational efficiency. On the GoPro and LOM datasets, integrating AIRS with different architectures resulted in overall improvements in PSNR and SSIM. On the GoPro dataset, AIRS consistently improved PSNR by approximately 0.2 to 0.4 dB, while also consistently improving SSIM. For example, when combined with DMPHN, AIRS improved PSNR by approximately 0.34 dB and significantly improved SSIM, indicating more reliable structure restoration in motion-blurred regions. Similarly, integrating AIRS into EVSSM further improved PSNR and SSIM values, demonstrating that selective information region modeling can be complementary even when combined with state-of-the-art deep deblurring architectures. On the LOM dataset, the performance improvements are typically more significant, especially in SSIM. For example, when AIRS is applied to GSAD and DiffLL, it improves PSNR by approximately 0.3–0.4 dB while achieving a stable SSIM boost. These results demonstrate that AIRS effectively suppresses noise amplification in low-information regions and enables the backbone network to focus on regions with structural information, thereby improving perception quality and maintaining structural consistency under complex low-light conditions.

[0060] Combination Figure 5 As shown, AIRS consistently reduces computational costs and inference time across different architectures. The integration of AIRS significantly reduces floating-point operations and runtime on both datasets without increasing model capacity. For example, on the GoPro dataset, combining AIRS with NAFNet reduces inference time by approximately 19% while simultaneously improving performance, demonstrating that AIRS effectively suppresses redundant computations in less important regions during feature modeling. Similarly, AIRS reduces GRL inference time by approximately 9%, indicating its compatibility with heavyweight architectures that rely on global modeling. The efficiency improvements are even more pronounced on the LOM dataset. For instance, AIRS significantly reduces inference time for DiffLL and PHP, which can be attributed to the large proportion of low-information background regions in low-light images. By selectively prioritizing informational regions, AIRS significantly reduces unnecessary computations in noise-dominated regions, enabling faster inference while maintaining or even improving recovery quality.

[0061] Experiments show that the method for restoring remote sensing images provided in this application can effectively achieve a good balance between restoration performance and computational efficiency.

[0062] The following describes the product embodiments provided in this application.

[0063] Combination Figure 6 As shown, this application provides an apparatus 600 for restoring remote sensing images. The apparatus 600 includes an image processing module 601, a region recognition module 602, and an image restoration module 603.

[0064] Image processing module 601 is used to extract features from degraded images to obtain feature maps of degraded images.

[0065] The region recognition module 602 is used to recognize the feature map of the degraded image in order to determine the key information region of the degraded image.

[0066] The image restoration module 603 is used to input the degraded image into the image restoration network model so that the image restoration network model can identify the key information regions of the degraded image and output the restored image.

[0067] Optionally, the region recognition module 602 is specifically used to: predict the probability that each feature vector in the feature map contains key information, thereby obtaining the classification result for each spatial location in the feature map. Based on the classification result for each spatial location in the feature map, the feature sequence of the feature map is semantically reordered to obtain the key information region of the degraded image.

[0068] Optionally, the region recognition module 602 is also used to: after determining the key information regions of the degraded image, assign corresponding state transition parameters to different key information regions in the degraded image.

[0069] Optionally, the image restoration module 603 is specifically used for: multi-scale encoding of each key information region of the degraded image to obtain multi-scale input features; weighted fusion of the multi-scale input features to obtain fused features; multi-scale decoding of the fused features to obtain image prediction results at different scales; and applying pixel-level supervision to the image prediction results at different scales to restore the degraded image and obtain the restored image.

[0070] The apparatus for restoring remote sensing images provided in this application is used to implement the method for restoring remote sensing images in the foregoing embodiments. The specific implementation methods and beneficial effects can be referred to the foregoing embodiments, and will not be repeated here.

[0071] Combination Figure 7As shown, this application provides an electronic device 700, including a processor 701 and a memory 702. Optionally, the electronic device 700 may further include a communication interface 703 and a bus 704. The processor 701, communication interface 703, and memory 702 can communicate with each other via the bus 704. The communication interface 703 can be used for information transmission. The processor 701 can call logical instructions in the memory 702 to execute the method for restoring remote sensing images described in the above embodiment.

[0072] Furthermore, the logic instructions in the aforementioned memory 702 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0073] The memory 702, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as program instructions / modules corresponding to the methods in the embodiments of this application. The processor 701 executes functional applications and data processing by running the program instructions / modules stored in the memory 702, that is, it implements the method for restoring remote sensing images in the above embodiments.

[0074] The memory 702 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 702 may include high-speed random access memory and may also include non-volatile memory.

[0075] This application provides a storage medium storing computer-executable instructions configured to perform the method for restoring remote sensing images described in the above embodiments.

[0076] The aforementioned storage medium can be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0077] The technical solutions of this application embodiment can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in this application embodiment. The aforementioned storage medium can be a non-transitory storage medium, including: USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, and other media capable of storing program code; it can also be a transient storage medium.

[0078] This application provides a computer program product, including a computer program that, when executed by a processor, implements the method for restoring remote sensing images described in the above embodiments.

[0079] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for describing embodiments only and is not intended to limit the claims. As used in the description of embodiments and claims, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used in this application means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, the relevant parts can be referred to the description of the method section.

[0080] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0081] The methods and products (including but not limited to devices and equipment) disclosed in the embodiments herein can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to implement this embodiment according to actual needs. In addition, the functional units in the embodiments of this application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0082] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description; sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

Claims

1. A method for recovering remotely sensed images, characterized in that, include: Feature extraction is performed on the degraded image to obtain the feature map of the degraded image; Identify the feature map of the degraded image to determine the key information region of the degraded image; The degraded image is input into an image restoration network model, which then identifies key information regions of the degraded image and outputs a restored image.

2. The method according to claim 1, characterized in that, The process of identifying the feature map of the degraded image to determine the key information regions of the degraded image includes: For each feature vector in the feature map, predict the probability that the feature vector contains key information to obtain the classification result for each spatial location in the feature map; Based on the classification results of each spatial location in the feature map, the feature sequence of the feature map is semantically reordered to obtain the key information region of the degraded image.

3. The method according to claim 1, characterized in that, After determining the key information region of the degraded image, the method further includes: Assign corresponding state transition parameters to different key information regions in the degraded image.

4. The method according to claim 1, characterized in that, The image restoration network model identifies key information regions of the degraded image and outputs a restored image, including: For each key information region of the degraded image, multi-scale encoding is performed on the key information regions to obtain multi-scale input features; The multi-scale input features are weighted and fused to obtain the fused features; The fused features are then decoded at multiple scales to obtain image prediction results at different scales. Pixel-level supervision is applied to the image prediction results at different scales to restore the degraded image, resulting in the restored image.

5. An apparatus for recovering remotely sensed images, characterized in that, include: An image processing module is used to extract features from a degraded image to obtain a feature map of the degraded image; A region recognition module is used to recognize the feature map of the degraded image in order to determine the key information region of the degraded image; The image restoration module is used to input the degraded image into the image restoration network model, so that the image restoration network model can identify the key information regions of the degraded image and output the restored image.

6. The apparatus according to claim 5, characterized in that, The region identification module is specifically used for: For each feature vector in the feature map, predict the probability that the feature vector contains key information to obtain the classification result for each spatial location in the feature map; Based on the classification results of each spatial location in the feature map, the feature sequence of the feature map is semantically reordered to obtain the key information region of the degraded image.

7. The apparatus according to claim 5, characterized in that, The region identification module is also used for: After determining the key information regions of the degraded image, corresponding state transition parameters are assigned to different key information regions in the degraded image.

8. The apparatus according to claim 5, characterized in that, The image restoration module is specifically used for: For each key information region of the degraded image, multi-scale encoding is performed on the key information regions to obtain multi-scale input features; The multi-scale input features are weighted and fused to obtain the fused features; The fused features are then decoded at multiple scales to obtain image prediction results at different scales. Pixel-level supervision is applied to the image prediction results at different scales to restore the degraded image, resulting in the restored image.

9. An electronic device comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to, when executing the program instructions, perform the method for recovering remotely sensed images as described in any one of claims 1 to 4.

10. A storage medium storing program instructions, characterized in that, When the program instructions are executed on a computer, they perform the method for restoring remotely sensed images as described in any one of claims 1 to 4.

Images