An anti-environment global image reconstruction method for single-machine perception information

Abstract

This invention discloses a global image reconstruction method for adversarial environments based on single-machine sensing information, belonging to the field of image reconstruction technology. The method includes preprocessing the original sensing image to obtain a preprocessed image; extracting local key features from the preprocessed image using a MaskAttn module and suppressing remaining interference features through an adaptive masking mechanism to output a high-quality local key feature matrix; performing global spatial alignment and semantic fusion on the local key feature matrix using an IMSAM module to extract a global feature matrix; deeply fusing the local key feature matrix and the global feature matrix; and generating a complete globally reconstructed image through a reconstruction generation module. This invention, through the collaborative work of the MaskAttn and IMSAM modules, balances local detail accuracy with global coherence, exhibits strong anti-interference capabilities, adapts to the real-time operation requirements of single-machine sensing devices, and can be widely applied to image reconstruction scenarios in single-machine sensing systems such as UAVs.

Description

Technical Field

[0001] This invention relates to the field of image reconstruction technology, and specifically to a global image reconstruction method for adversarial environments based on single-machine perception information. Background Technology

[0002] With the rapid popularization of unmanned aerial vehicle (UAV) systems, stand-alone sensing technology has been widely applied in many fields such as environmental monitoring, emergency rescue, intelligent navigation, and field exploration. One of the core functions of a stand-alone sensing system is to acquire image information of the surrounding environment through its onboard image sensors and to reconstruct a global image based on this information, providing reliable visual support for subsequent tasks such as target recognition, path planning, and environmental understanding. The core requirement of global image reconstruction is to integrate the fragmented and localized image data acquired by the stand-alone system into a complete, coherent, and clear global image, while also dealing with the inevitable environmental interference in practical applications.

[0003] Adversarial environments refer to various interference factors encountered during single-device perception, mainly including electromagnetic interference, sudden changes in light, obstacle occlusion, noise pollution, and image distortion. These interferences can cause problems such as blurring, missing elements, artifacts, and pixel distortion in the original perceived images acquired by the single-device, severely affecting the effectiveness of global image reconstruction. Currently, global image reconstruction methods for UAV perception information are mainly divided into two categories: traditional reconstruction methods and deep learning-based reconstruction methods. Traditional reconstruction methods, such as interpolation, Bayesian estimation, and stitching fusion, have low computational complexity and do not require high computing power from the UAV, but they have weak anti-interference capabilities. In adversarial environments, they are difficult to effectively distinguish between useful features and interference features, and the reconstructed global images often suffer from problems such as loss of detail, poor global coherence, and obvious artifacts, failing to meet the needs of practical applications.

[0004] Existing image reconstruction methods mostly focus on improving reconstruction accuracy, ignoring the resource constraints of UAV equipment. The model structure is complex and the computational load is large, making it difficult to achieve real-time reconstruction on a single machine. At the same time, existing methods have not achieved deep collaboration between local key feature extraction and global feature alignment, and cannot effectively solve the core problems of local feature interference and global information fragmentation in adversarial environments.

[0005] To address the aforementioned issues, there is an urgent need for a global image reconstruction method for adversarial environments that is based on single-machine perception information, in order to solve the problems existing in traditional methods. Summary of the Invention

[0006] The purpose of this invention is to provide a global image reconstruction method for adversarial environments based on single-machine perception information. By deeply integrating the Mask Attention CNN module (MaskAttn) with the improved MobileSAM (IMSAM), it achieves accurate extraction and interference suppression of local key features and efficient alignment of global features, taking into account the resource constraints of UAV equipment and the reconstruction effect in adversarial environments, thereby improving the robustness, coherence and detail accuracy of global image reconstruction.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for global image reconstruction in adversarial environments based on single-machine perception information includes: Step 1: Acquire raw sensing images based on the UAV's sensing equipment and preprocess them to obtain preprocessed images; Step 2: Extract local key features from the preprocessed image based on the MaskAttn module, and suppress the remaining interference features through an adaptive masking mechanism to output a high-quality local key feature matrix; Step 3: Based on the IMSAM module, perform global spatial alignment and semantic fusion on the local key feature matrices to extract the global feature matrix; Step 4: Deeply fuse the local key feature matrix and the global feature matrix, and generate a complete global reconstructed image through the reconstruction generation module.

[0008] Furthermore, in step 1, the preprocessing is a normalization process.

[0009] Furthermore, in step 2, the MaskAttn module includes a basic convolutional layer, an attention calculation layer, a mask generation layer, a feature selection layer, and a feature fusion layer.

[0010] Furthermore, in step 2, local key features of the preprocessed image are extracted based on the MaskAttn module, and the remaining interfering features are suppressed through an adaptive masking mechanism to output a high-quality local key feature matrix, specifically: Step 201: Initialize the MaskAttn module; Step 202: Input the preprocessed image into the basic convolutional layer of the MaskAttn module, extract the basic local features of the image through lightweight two-dimensional convolution operation, and output the basic local feature matrix; Step 203: Input the basic local feature matrix into the attention calculation layer of the MaskAttn module to calculate the attention weight matrix; Step 204: Input the attention weight matrix into the mask generation layer of the MaskAttn module to generate an adaptive mask matrix; Step 205: Input the basic local feature matrix and the adaptive mask matrix into the feature filtering layer of the MaskAttn module, perform element-wise multiplication, and obtain the filtered local key feature matrix. Step 206: Input the filtered local key feature matrix into the feature fusion layer of the MaskAttn module. The features are integrated through a single two-dimensional convolution operation to eliminate local feature gaps generated during feature filtering. At the same time, the fused features are batch normalized to obtain a high-quality local key feature matrix.

[0011] Furthermore, in step 3, the IMSAM module includes an optimized Tiny-ViT encoder, an optimized decoder, and a global feature correction layer.

[0012] Furthermore, in step 3, the local key feature matrices are globally spatially aligned and semantically fused based on the IMSAM module to extract the global feature matrix, specifically as follows: Step 301: Initialize the IMSAM module; Step 302: Input the local key feature matrix into the Tiny-ViT encoder of the IMSAM module. Through the multi-layer attention mechanism and convolution operation of the Tiny-ViT encoder, the local key features are encoded, global feature information is extracted, and the encoded global feature matrix is ​​output. Step 303: Input the encoded global feature matrix into the optimized decoder of the IMSAM module. The optimized decoder performs global spatial alignment on the encoded global feature matrix through a global attention mechanism, corrects the spatial position deviation of local features, and obtains global encoded features. At the same time, the optimized decoder receives local key feature matrices through the feature fusion interface, performs preliminary fusion of local key feature matrices with global encoded features, and outputs the decoded global feature matrix. Step 304: Input the decoded global feature matrix into the global feature correction layer of the IMSAM module. Through a lightweight feature correction algorithm, fine-tune the pixel values ​​and spatial positions of the global features to correct the deviations generated during the global feature alignment process. Step 305: Fuse the corrected global feature matrix with the local key feature matrix to obtain and output the final global feature matrix.

[0013] Furthermore, the reconstruction generation module adopts a lightweight transposed convolutional network structure.

[0014] Furthermore, in step 4, the local key feature matrix and the global feature matrix are deeply fused, and a complete global reconstructed image is generated through the reconstruction generation module, specifically as follows: Step 401: Combine the local key feature matrix and the global feature matrix by addition; Step 402: Initialize the refactoring generation module; Step 403: Input the feature matrix after addition and fusion into the reconstruction generation module, perform upsampling through two-dimensional transpose convolution operation, gradually restore the size and details of the image, and output the preliminary global reconstructed image; Step 404: Post-process and optimize the preliminary global reconstructed image by using a pixel smoothing algorithm to eliminate jagged artifacts generated during the reconstruction process, and obtain the final global reconstructed image.

[0015] Furthermore, the method also includes: Step 5: Iteratively update the parameters of the MaskAttn module, IMSAM module, and reconstruction generation module based on the reconstruction error.

[0016] In summary, the present invention has at least one of the following beneficial technical effects: 1. This invention utilizes the adaptive mask attention mechanism of the MaskAttn module to accurately extract key local features and suppress interference features in adversarial environments. At the same time, the IMSAM module optimizes the anti-interference feature extraction capability. The two work together to effectively solve problems such as reconstruction distortion and artifacts caused by interference in adversarial environments, thereby improving the robustness of the reconstruction method.

[0017] 2. This invention extracts local key features through the MaskAttn module and aligns global features through the IMSAM module. The deep integration of the two takes into account both the accuracy of local details and the global coherence of the reconstructed image. It solves the problems of discontinuous global reconstruction and loss of details caused by fragmentation of single-machine perception images. It achieves an organic combination of local feature extraction and interference suppression, and global feature alignment. It is suitable for adversarial environments and single-machine requirements, and breaks through the technical bottlenecks of insufficient coordination between local and global features and difficulty in balancing anti-interference and lightweight in existing methods. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0020] like Figure 1 As shown, this invention provides a method for global image reconstruction in adversarial environments based on single-machine sensing information, including: Step 1: Acquire raw sensing images based on the UAV's sensing equipment and preprocess them to obtain preprocessed images; Step 2: Extract local key features from the preprocessed image based on the MaskAttn module, and suppress the remaining interference features through an adaptive masking mechanism to output a high-quality local key feature matrix; Step 3: Based on the IMSAM module, perform global spatial alignment and semantic fusion on the local key feature matrices to extract the global feature matrix; Step 4: Deeply fuse the local key feature matrix and the global feature matrix, and generate a complete global reconstructed image through the reconstruction generation module.

[0021] In step 1, the raw sensing images are acquired based on the UAV's sensing equipment and then preprocessed to obtain preprocessed images, specifically as follows: For the raw sensor images acquired by the UAV's sensing equipment, the image pixel values ​​are normalized to the [0,1] interval to eliminate the influence of differences in pixel value magnitude. The normalization formula is: ; In the formula, These are the original pixel values. The minimum pixel value in the image. The maximum pixel value in the image. The normalized pixel values; keeping the pixel positions unchanged: all calculated values... By strictly arranging the pixels according to the original image, a standardized preprocessed image is obtained, denoted as... This serves as the input for subsequent adversarial interference suppression and feature extraction.

[0022] In step 2, the MaskAttn module includes a basic convolutional layer, an attention calculation layer, a mask generation layer, a feature selection layer, and a feature fusion layer. The MaskAttn module first initializes the parameters, then extracts basic local features of the image through the basic convolutional layer, calculates attention weights through the attention calculation layer, generates an adaptive mask matrix through the mask generation layer, selects useful local features through the feature selection layer, and finally integrates and standardizes the features through the feature fusion layer to output a local key feature matrix.

[0023] In step 2, local key features of the preprocessed image are extracted based on the MaskAttn module, and the remaining interference features are suppressed through an adaptive masking mechanism to output a high-quality local key feature matrix, specifically: Step 201: Initialize the MaskAttn module by initializing the parameters of each layer, including the convolution kernels and bias terms of the basic convolutional layers, the convolution kernels and bias terms of the attention calculation layers, and the adaptive threshold of the mask generation layer. All parameters are initialized to random small values ​​and subsequently updated iteratively through model optimization. The basic convolutional layer adopts a lightweight design, and the number of convolutional kernels is adaptively set according to the drone's computing power to avoid excessive computation. Step 202: Basic Local Feature Extraction: Extracting the preprocessed image The input is fed into the base convolutional layer of the MaskAttn module, which extracts basic local features of the image through lightweight 2D convolution operations and outputs a basic local feature matrix. The stride is determined by the number of convolutional kernels in the basic convolutional layer. The convolutional operation in the basic convolutional layer uses a stride of 1 to ensure that the size of the feature matrix is ​​consistent with the input image and to avoid loss of feature size. Step 203: Attention Weight Calculation: Calculate the basic local feature matrix. The input is fed into the attention computation layer to calculate the attention weight matrix. The specific formula is as follows: ; in, It is the attention weight matrix. It is a two-dimensional convolution operation. It is the basic local feature matrix output by the basic convolutional layer. It is the convolution bias term; Using two-dimensional convolution kernels For the basic local feature matrix Perform convolutional transformations to extract feature correlations to support attention weight calculations. The convolution kernel for two-dimensional convolution operations. A 3×3 square convolution kernel is used to balance feature correlation extraction and computational efficiency. The activation function normalizes the attention weights to the [0,1] interval, ensuring that the weights of each feature pixel accurately reflect its usefulness; Step 204: Adaptive Mask Generation: Generate the attention weight matrix The input is fed into the mask generation layer to generate an adaptive mask matrix. The specific formula is as follows: ; in, It is a threshold function. Attention weight matrix Using adaptive threshold Attention weight matrix A threshold is applied to generate a binary mask matrix that filters useful features and suppresses interfering features. When the attention weights... When a feature is deemed useful, the corresponding position in the mask matrix is ​​set to 1, and the feature is retained; when the attention weights... When this is detected, it is determined to be an interference feature, and the corresponding position in the mask matrix is ​​set to 0 to suppress the feature; Step 205: Local Feature Filtering: Filter the basic local feature matrix. With adaptive mask matrix Perform element-wise multiplication to obtain the filtered local key feature matrix. This matrix retains only useful features with high attention weights from the basic features, while suppressing interfering features with low weights. The specific formula is as follows: ; in, It is an element-wise product operation; Step 206: Feature Fusion and Standardization: The filtered local key feature matrix... The input is fed into the feature fusion layer, where features are integrated through a single 2D convolution operation to eliminate local feature gaps generated during feature selection. At the same time, the fused features are batch normalized to accelerate subsequent model training and improve feature stability.

[0024] In step 3, the IMSAM module is an improvement on the traditional MobileSAM, including an optimized Tiny-ViT encoder, an optimized decoder, and a global feature correction layer. The IMSAM module encodes the local key feature matrix, decodes the global spatial alignment, corrects the features, and fuses local and global features to output a global feature matrix, thus solving the problem of insufficient global coherence caused by fragmentation of single-machine sensing images.

[0025] In step 3, the local key feature matrices are globally spatially aligned and semantically fused based on the IMSAM module to extract the global feature matrix, specifically as follows: Step 301: Initialize the IMSAM module. The IMSAM module is based on the traditional MobileSAM, with improvements and parameter initialization. Specifically, the optimized Tiny-ViT encoder retains the core Tiny-ViT structure of the traditional MobileSAM, removes 2-3 redundant convolutional layers to reduce computation, and adjusts the number and size of convolutional kernels to improve anti-interference feature extraction capabilities. The optimized decoder incorporates a feature fusion interface to adapt to local key feature matrices. The input; the global feature correction layer uses a lightweight feature correction algorithm to initialize the correction parameters; Step 302: Local Feature Encoding: Encode the local key feature matrix The optimized Tiny-ViT encoder, input to the IMSAM module, encodes local key features through multi-layer attention mechanisms and convolutional operations, extracts global feature information, and outputs the encoded global feature matrix. The optimized Tiny-ViT encoder adopts a block-based attention calculation method, which divides the feature matrix into blocks and performs attention calculations to reduce computational redundancy, improve coding efficiency, and adapt to the computing power of drones. Step 303: Global Feature Alignment and Decoding: Align the encoded global feature matrix... The input is fed into an optimized decoder, which uses a global attention mechanism to... Global spatial alignment is performed to correct spatial positional deviations of local features, ensuring the consistency of different local features in the global space. Simultaneously, the decoder receives local key feature matrices through a feature fusion interface. The local features and global encoded features are initially fused to output the decoded global feature matrix. ; Step 304: Global Feature Correction: Correct the decoded global feature matrix The input is fed into the global feature correction layer. A lightweight feature correction algorithm is used to fine-tune the pixel values ​​and spatial positions of the global features, correcting the deviations generated during the global feature alignment process. The feature correction formula is as follows: ; in, This is the corrected global feature matrix. This is the decoded global feature matrix. For correction factors, The value is 0.6. To correct the bias term, a value of 0.02 is used; It is an optimized decoder for the IMSAM module. It is an optimized Tiny-ViT encoder for the IMSAM module; Step 305: Local and global feature fusion: The corrected global feature matrix is ​​then fused... With local key feature matrix The fusion process is performed using the following formula: ; in, It is the global feature matrix output by the IMSAM module. It is an optimized Tiny-ViT encoder based on the IMSAM module, which uses feature fusion coefficients. Adjusting the weights of local features enables deep collaboration between local and global features, ultimately outputting the global feature matrix. ,in, The value is 0.3.

[0026] The reconstruction generation module adopts a lightweight transposed convolutional network structure to upsample the fused feature matrix, gradually restore the image size and details, and output a preliminary global reconstructed image.

[0027] In step 4, the local key feature matrix and the global feature matrix are deeply fused, and a complete global reconstructed image is generated through the reconstruction generation module. Specifically: Step 401: Feature Fusion: According to The method of additive fusion combines the local key feature matrix with the global feature matrix to achieve an organic combination of local detailed features and global coherent features. The additive fusion method has low computational complexity, is compatible with the computing power of UAVs, and can effectively retain the feature advantages of both, avoiding the computational redundancy caused by complex fusion methods. Step 402: Initialization of the reconstruction generation module: The reconstruction generation module adopts a lightweight transposed convolutional network structure, consisting of 2-3 two-dimensional transposed convolutional layers. The parameters such as the convolutional kernel and bias term of the transposed convolutional layer are initialized. The convolutional kernel size of the transposed convolutional layer is 3×3, and the stride is set according to the size ratio of the feature matrix and the image to ensure that the fused feature matrix can be upsampled to the original image size. Step 403: Global Image Reconstruction: Reconstruct the fused feature matrix The image is input to the reconstruction generation module, where it is upsampled through a 2D transposed convolution operation to gradually restore the image's size and details, outputting a preliminary global reconstructed image Y. The specific formula is as follows: ; in, The feature matrix after fusing local and global data. Two-dimensional transposed convolution upsampling is performed to restore the image size and achieve global image reconstruction. It is a two-dimensional transposed convolution operation. The ReLU activation function is used during the transposed convolution process to improve the detail representation of the reconstructed image and avoid pixel value distortion. Step 404: Reconstructed Image Optimization: Post-processing optimization is performed on the initial global reconstructed image Y. A pixel smoothing algorithm is used to eliminate jagged artifacts generated during the reconstruction process. The pixel smoothing formula is as follows: ; in, To optimize the image The pixel value of the location, for The total number of pixels in the location's neighborhood. for The set of neighboring pixels of a location. To initially reconstruct the image in The pixel values ​​of the location; at the same time, adjust the contrast and brightness of the image to make it consistent with the visual effect of the original, undisturbed image; Step 405: Output the final global reconstructed image: After post-processing optimization, the final global reconstructed image is obtained. The imagery balances local detail accuracy with global coherence, exhibits strong anti-interference capabilities, and can meet the practical application requirements of stand-alone sensing systems.

[0028] The method further includes: Step 5: Iteratively update the parameters of the MaskAttn module, IMSAM module, and reconstruction generation module based on the reconstruction error.

[0029] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0030] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0031] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0032] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0033] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A method for global image reconstruction in adversarial environments based on single-machine sensing information, characterized in that, include: Step 1: Acquire raw sensing images based on the UAV's sensing equipment and preprocess them to obtain preprocessed images; Step 2: Extract local key features from the preprocessed image based on the MaskAttn module, and suppress the remaining interference features through an adaptive masking mechanism to output a high-quality local key feature matrix; Step 3: Based on the IMSAM module, perform global spatial alignment and semantic fusion on the local key feature matrices to extract the global feature matrix; Step 4: Deeply fuse the local key feature matrix and the global feature matrix, and generate a complete global reconstructed image through the reconstruction generation module.

2. The method for global image reconstruction in adversarial environments based on single-machine sensing information according to claim 1, characterized in that, In step 1, the preprocessing is a normalization process.

3. The method for global image reconstruction in adversarial environments based on single-machine sensing information according to claim 1, characterized in that, In step 2, the MaskAttn module includes a basic convolutional layer, an attention calculation layer, a mask generation layer, a feature selection layer, and a feature fusion layer.

4. The method for global image reconstruction in adversarial environments based on single-machine sensing information according to claim 3, characterized in that, In step 2, local key features of the preprocessed image are extracted based on the MaskAttn module, and the remaining interference features are suppressed through an adaptive masking mechanism to output a high-quality local key feature matrix, specifically: Step 201: Initialize the MaskAttn module; Step 202: Input the preprocessed image into the basic convolutional layer of the MaskAttn module, extract the basic local features of the image through lightweight two-dimensional convolution operation, and output the basic local feature matrix; Step 203: Input the basic local feature matrix into the attention calculation layer of the MaskAttn module to calculate the attention weight matrix; Step 204: Input the attention weight matrix into the mask generation layer of the MaskAttn module to generate an adaptive mask matrix; Step 205: Input the basic local feature matrix and the adaptive mask matrix into the feature filtering layer of the MaskAttn module, perform element-wise multiplication, and obtain the filtered local key feature matrix. Step 206: Input the filtered local key feature matrix into the feature fusion layer of the MaskAttn module. The features are integrated through a single two-dimensional convolution operation to eliminate local feature gaps generated during feature filtering. At the same time, the fused features are batch normalized to obtain a high-quality local key feature matrix.

5. The method for global image reconstruction in adversarial environments based on single-machine sensing information according to claim 4, characterized in that, In step 3, the IMSAM module includes an optimized Tiny-ViT encoder, an optimized decoder, and a global feature correction layer.

6. A method for global image reconstruction in adversarial environments based on single-machine sensing information, as described in claim 5, is characterized in that... In step 3, the local key feature matrices are globally spatially aligned and semantically fused based on the IMSAM module to extract the global feature matrix, specifically as follows: Step 301: Initialize the IMSAM module; Step 302: Input the local key feature matrix into the Tiny-ViT encoder of the IMSAM module. Through the multi-layer attention mechanism and convolution operation of the Tiny-ViT encoder, the local key features are encoded, global feature information is extracted, and the encoded global feature matrix is ​​output. Step 303: Input the encoded global feature matrix into the optimized decoder of the IMSAM module. The optimized decoder performs global spatial alignment on the encoded global feature matrix through a global attention mechanism, corrects the spatial position deviation of local features, and obtains global encoded features. At the same time, the optimized decoder receives local key feature matrices through the feature fusion interface, performs preliminary fusion of local key feature matrices with global encoded features, and outputs the decoded global feature matrix. Step 304: Input the decoded global feature matrix into the global feature correction layer of the IMSAM module. Through a lightweight feature correction algorithm, fine-tune the pixel values ​​and spatial positions of the global features to correct the deviations generated during the global feature alignment process. Step 305: Fuse the corrected global feature matrix with the local key feature matrix to obtain and output the final global feature matrix.

7. A method for global image reconstruction in adversarial environments based on single-machine sensing information according to claim 6, characterized in that, The reconstruction generation module adopts a lightweight transposed convolutional network structure.

8. A method for global image reconstruction in adversarial environments based on single-machine sensing information according to claim 7, characterized in that, In step 4, the local key feature matrix and the global feature matrix are deeply fused, and a complete global reconstructed image is generated through the reconstruction generation module. Specifically: Step 401: Combine the local key feature matrix and the global feature matrix by addition; Step 402: Initialize the refactoring generation module; Step 403: Input the feature matrix after addition and fusion into the reconstruction generation module, perform upsampling through two-dimensional transpose convolution operation, gradually restore the size and details of the image, and output the preliminary global reconstructed image; Step 404: Post-process and optimize the preliminary global reconstructed image by using a pixel smoothing algorithm to eliminate jagged artifacts generated during the reconstruction process, and obtain the final global reconstructed image.

9. A method for global image reconstruction in adversarial environments based on single-machine sensing information according to claim 1, characterized in that, The method further includes: Step 5: Iteratively update the parameters of the MaskAttn module, IMSAM module, and reconstruction generation module based on the reconstruction error.

Images