A method and system for multi-resolution collaborative real-time deblurring of unmanned aerial vehicle aerial images

Abstract

The application provides a kind of unmanned aerial vehicle aerial image multi-resolution cooperative real-time deblurring method and system, belongs to image processing technical field, comprising: shallow feature is extracted to high, medium, low resolution image of unmanned aerial vehicle aerial photography, after being aligned size by upsampling, cascade, obtain high, medium resolution fusion feature.High, medium cascade feature and low resolution shallow feature are summed with original image element by element, and cooperative guide feature is obtained.On each branch, feature is processed by convolution, compression and inflation of different depth, residual connection is aggregated as multi-receptive field representation.After pixel attention weighting, medium, low resolution weighted feature is upsampled and cascaded with high resolution feature, and then added with original high resolution image to reconstruct clear image.The method effectively handles non-uniform blur caused by multi-target and irregular motion of unmanned aerial vehicle through low resolution guidance, cross-resolution feature association and multi-receptive field fusion, and restores clear aerial image.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method and system for multi-resolution collaborative real-time deblurring of drone aerial images. Background Technology

[0002] Drone aerial imagery, captured by cameras mounted on drones, produces still images or dynamic video frames from the air, and has become a crucial support for high-quality development in fields such as scene recognition, identity verification, and data acquisition. However, the dynamic flexibility of drones brings with it the problem of image blurring, especially during high-speed flight or attitude adjustments. Drones are susceptible to camera shake, atmospheric disturbances, and sensor limitations, leading to decreased image quality and hindering their effectiveness in precise identification and long-distance monitoring.

[0003] Recent breakthroughs in deep learning technology have injected new momentum into the field of image deblurring, giving rise to a series of data-driven end-to-end solutions. Deep learning-based methods employ convolutional neural networks, generative adversarial networks, or recurrent neural networks to model the mapping from blurred to sharp images. However, existing methods generally neglect feature associations between different resolutions, making it difficult to fully capture multi-scale blur cues caused by rapid pose changes, resulting in insufficient adaptability of the models to non-uniform blur. Secondly, most existing technologies employ extremely deep network architectures for training, leading to high computational latency and slow inference speed, making them unsuitable for time-sensitive scenarios. Summary of the Invention

[0004] This invention addresses the technical problems in the prior art by providing a multi-resolution collaborative real-time deblurring method for UAV aerial images. This method can solve the problem of blurring in aerial images caused by multi-target motion or irregular movement of UAVs.

[0005] Includes the following steps: Acquire high-resolution, medium-resolution, and low-resolution images from drone aerial photography; Extract shallow features from images at three different resolutions; High-resolution features are obtained by upsampling medium-resolution and low-resolution shallow features to the same size as high-resolution shallow features and then concatenating them; medium-resolution features are obtained by downsampling and upsampling high-resolution and low-resolution shallow features to the same size as medium-resolution shallow features, respectively. High-resolution concatenated features are obtained by upsampling medium-resolution features and concatenating them with high-resolution features; medium-resolution concatenated features are obtained by upsampling low-resolution shallow features and concatenating them with medium-resolution features; multi-resolution collaborative guidance features are obtained by summing the high-resolution concatenated features, medium-resolution concatenated features, and low-resolution shallow features element-wise with the original resolution map. The multi-resolution collaborative guided features are sequentially extracted at different depths, including convolution, compression and dilation. The features at different depths are then aggregated through residual connections to obtain a multi-receptive field aggregated representation. The multi-receptive field aggregation representation is weighted by pixel attention, and after upsampling the weighted features at medium and low resolutions, the three resolution features are concatenated and summed element-wise with the high-resolution image to obtain the reconstructed high-resolution deblurred image.

[0006] Preferably, the multi-resolution collaborative guidance features are obtained through processing by a multi-resolution collaborative module, which includes three feature processing branches with different resolutions; the shallow feature extraction in the feature processing branches adopts a shallow convolution module. In this process, the high-resolution and medium-resolution feature processing branches are convolutional after cascading to obtain high-resolution and medium-resolution features; each feature processing branch undergoes a convolutional layer before element-wise summation with the original resolution image. The shallow convolution module employs... Convolution and Convolutional structure; high-resolution and medium-resolution feature processing branches are processed in convolution after cascading. and The convolution.

[0007] Preferably, the medium-resolution and low-resolution UAV aerial images are obtained by sequentially convolving and downsampling the high-resolution images.

[0008] Preferably, the multi-receptive-field aggregation representation is implemented through a feature aggregation module; the feature aggregation module is implemented through... The convolution process is applied to the collaboratively guided features, through convolution in the channel direction. Compression and expansion are performed to obtain deep feature extraction.

[0009] Preferably, the multi-receptive-field aggregation represents a reconstruction module that performs pixel attention weighting and adopts a pixel attention mechanism; The reconstruction module of the pixel attention mechanism uses two Convolution, LeakyReLU function and sigmoid function are used for attention weighting, and the correlation between pixels is fused through weight matrix to obtain weighted features.

[0010] Preferably, the upsampling of the medium and low resolution weighted features employs a deconvolution layer, wherein the deconvolution layer uses a set of deconvolution filters to upsample the features.

[0011] Preferably, after performing pixel attention weighting on the multi-receptive field aggregation representation, the weighted features after upsampling at medium and low resolution are summed element-wise with the original medium and low resolution images to obtain the reconstructed medium and low resolution deblurred images. Specifically, the average L1 norm error is constructed based on the multi-resolution image reconstruction loss of the training set and used as the loss function for training.

[0012] This invention also proposes a multi-resolution collaborative real-time deblurring system for UAV aerial images, the system comprising: The data acquisition unit is used to acquire high-resolution, medium-resolution, and low-resolution images taken by the UAV. The multi-resolution collaborative guidance unit is used to extract shallow features from images of three resolutions; it is also used to upsample the shallow features of medium and low resolutions to the same size as the shallow features of high resolution, and then concatenate them to obtain high-resolution features; to downsample and upsample the shallow features of high resolution and low resolutions to the same size as the shallow features of medium resolution, and then concatenate them to obtain medium-resolution features; it is also used to upsample the medium-resolution features and concatenate them with the high-resolution features to obtain high-resolution concatenated features; to upsample the shallow features of low resolutions and concatenate them with the medium-resolution features to obtain medium-resolution concatenated features; and to sum the high-resolution concatenated features, medium-resolution concatenated features, and low-resolution shallow features element-wise with the original resolution image to obtain multi-resolution collaborative guidance features. The multi-receptive-field aggregation unit is used to sequentially extract features of different depths from multi-resolution collaboratively guided features, including convolution, compression and dilation, and aggregate the features of different depths through residual connections to obtain the multi-receptive-field aggregated representation. The deblurring image processing unit is used to perform pixel attention weighting on the multi-receptive field aggregation representation, and after upsampling the medium and low resolution weighted features, the three resolution features are concatenated and summed element-wise with the high resolution image to obtain the reconstructed high resolution deblurring image.

[0013] The present invention also proposes a computer-readable storage medium storing a data processing program, which, when executed by a processor, implements the aforementioned multi-resolution collaborative real-time deblurring method for UAV aerial images.

[0014] The present invention also proposes a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned multi-resolution collaborative real-time deblurring method for UAV aerial images.

[0015] The beneficial effects of this invention are: This invention proposes a multi-resolution collaborative real-time deblurring method for UAV aerial images. This method utilizes a deep model to autonomously learn deep features in blurred images, avoiding explicit estimation of the blur kernel. To enable stable, clear, and high-detail real-time visual perception for UAVs during high-speed movement, this method achieves synergistic optimization of accuracy and speed through end-to-end real-time model processing. This makes it better suited for time-sensitive complex scenarios and supports the effective implementation of subsequent high-level visual tasks such as small target recognition and fine object classification. This invention employs multi-resolution branches to collaboratively process motion blur. It primarily guides the high-resolution processing process with key information (obtained through convolutional downsampling) from low resolution (processed by an attention mechanism). Compared to processing at only a single resolution, this method involves more parameters but is more robust, adapting to different resolution variations. Furthermore, combining multi-resolution branch processing with an attention mechanism offers significant advantages in computational cost compared to simple convolution operations, making it easier to obtain real-time processing results and potentially allowing for more flexible hardware requirements. Attached Figure Description

[0016] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the multi-resolution collaborative guidance module according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the feature aggregation module according to an embodiment of the present invention; Figure 4 This is the convolution in the channel direction of this embodiment of the invention. A schematic diagram of the structure; Figure 5 This is a schematic diagram of the reconstruction module of the pixel attention mechanism in an embodiment of the present invention. Detailed Implementation

[0017] 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.

[0018] Example 1 This embodiment proposes a multi-resolution collaborative real-time deblurring method for UAV aerial images. The main steps are as follows: Figure 1 As shown, the method includes: S1: Acquire high-resolution, medium-resolution, and low-resolution images taken by the drone.

[0019] S2: Extract shallow features from images of three resolutions; upsample the medium and low resolution shallow features to the same size as the high resolution shallow features, and cascade them to obtain high resolution features; downsample and upsample the high resolution and low resolution shallow features to the same size as the medium resolution shallow features, and cascade them to obtain medium resolution features.

[0020] High-resolution cascaded features are obtained by upsampling medium-resolution features and concatenating them with high-resolution features; medium-resolution cascaded features are obtained by upsampling low-resolution shallow features and concatenating them with medium-resolution features; and multi-resolution collaborative guidance features are obtained by summing the high-resolution cascaded features, medium-resolution cascaded features, and low-resolution shallow features element-wise with the original resolution map.

[0021] S3: Extract features of different depths sequentially from the multi-resolution collaborative guided features, including convolution, compression and dilation. Aggregate the features of different depths through residual connections to obtain a multi-receptive field aggregated representation.

[0022] S4: The multi-receptive field aggregation representation is weighted by pixel attention, and after upsampling the medium and low resolution weighted features, the three resolution features are concatenated and summed element-wise with the high resolution image to obtain the reconstructed high resolution deblurred image.

[0023] Specifically, aerial blur is mainly caused by the high-speed movement of the vehicle and lens shake, which includes both modelable global motion trajectories and unpredictable local high-frequency jitter. Images at different resolutions... It contains rich scene structure and motion information, among which Representing the Aerial images at a resolution of [resolution value] This represents the number of resolutions for aerial images, and the sizes of aerial images at different resolutions are as follows: .like Figure 2 As shown, the model framework in this embodiment adopts... Three different resolutions are used as inputs, with the low-resolution input obtained by downsampling the high-resolution input. Downsampling is performed using convolution, which, compared to fixed-step sampling, mean sampling, or pooling layers (which are fixed degradation methods), can model a richer degradation process from high resolution to low resolution.

[0024] The entire model of this invention consists of a multi-resolution collaborative guidance module, a feature aggregation module, and a pixel attention mechanism reconstruction module, which are composed of three branches.

[0025] The multi-resolution collaborative boot module includes three resolution branches, with the high-resolution and medium-resolution branches including shallow modules. Convolution and convolution), and The convolutional layers, and Convolutional layers; low-resolution branches include shallow modules ( Convolution and (convolution) and The convolutional layers are used. In one layer, the features from the medium and low resolution branches are processed in the shallow module and then upsampled to the same size as the high-resolution features. Similarly, the features from the high and low resolution branches are processed in the shallow module and then downsampled and upsampled to the same size as the medium-resolution features. The features from both branches are then concatenated, and finally, the features from both branches are processed... and Convolutions yield high-resolution features. and medium resolution features Low- and medium-resolution features are convolved and upsampled with high-resolution features respectively. and medium resolution features After concatenating medium- and high-resolution features, each branch is then processed through a convolutional layer and summed element-wise with the original image to obtain multi-resolution collaborative guided features for each branch. The feature aggregation model consists of two sets. convolutional layers and Convolutional layers; residual connections between two sets of convolutional layers. Each branch in the reconstruction module of the pixel attention mechanism includes two... Convolutional, LeakyReLU, and Sigmoid functions are used for attention weighting. Low- and medium-resolution branches are then connected to deconvolutional layers, and high-resolution branches are then connected to... The convolutional layers; where the high-resolution branches are in Before convolutional compression, features from other resolutions are concatenated. Finally, the outputs of each branch of the pixel attention mechanism's reconstruction module are summed element-wise with the input image to obtain the reconstructed image.

[0026] Specifically, the acquisition of multi-resolution collaborative guidance features in S2 employs a multi-resolution collaborative guidance module. A schematic diagram of the specific model is shown below. Figure 2 As shown.

[0027] In processing images of different resolutions, the most basic features are first extracted using a shallow convolutional module. This shallow module employs... Convolution and Convolutional structures, with smaller kernels, retain more detailed information, which facilitates subsequent interaction between information of different resolutions. Since features of different resolutions have good spatial consistency, they can guide each other during image degradation modeling. Information in low-resolution features is strongly correlated with and more generalized than the corresponding high-resolution information in space, while information in high-resolution features is more specific. Features of different resolutions are fused through two paths: on one hand, the medium and low-resolution features are upsampled to the same size as the high-resolution features; on the other hand, the high and low-resolution features are downsampled and upsampled to the same size as the medium-resolution features. The features on both branches are then concatenated, and the features on both branches are then processed... and Convolutions yield high-resolution features. and medium resolution features Low- and medium-resolution features are concatenated with medium- and high-resolution features after convolution and upsampling, respectively. Then, after passing through a convolution layer on their respective resolution processing branches, they are summed element-wise with the original image to obtain the output guided by the three resolutions. ,in 3}.

[0028] Specifically, the feature aggregation model in S3 is illustrated in the diagram below. Figure 3 As shown.

[0029] This invention also focuses on exploring the crucial role of multi-receptive field collaborative analysis in motion blur feature extraction. Because motion blur exhibits significant scale variability and directional diversity in image space, a single receptive field is insufficient to fully capture its complex degradation patterns. By constructing a multi-receptive field collaborative analysis architecture, hierarchical analysis and reconstruction of motion blur features can be achieved. The feature aggregation module primarily mines key information within a single resolution. First, shallow feature channels typically correspond to fine-grained detail information (such as edges and textures). These features are extremely sensitive to slight blurring; a decrease in their activity or a change in their pattern directly reflects the degree of loss of high-frequency information and the local characteristics of key objects. Furthermore, mid-level feature channels carry more complex structural information. Changes in the response of these channel features help characterize the continuity and directionality of blurred regions, thereby modeling and describing the macroscopic trend of motion blur. In addition, deep feature channels correspond to global, semantic-level contextual information. The activation patterns they present are more abstract and richer, revealing the overall and differentiated patterns exhibited by different types of scenes (such as cities, farmland, and water bodies) after undergoing motion blur degradation. This global representation at the channel level provides key clues for understanding the interaction between fuzziness and scene content.

[0030] To fully extract this information, the feature aggregation module further processes features at different resolutions. This module captures features from different receptive fields as the network deepens, then uses soft attention to enhance key information in each channel based on these features, and mines the correlations between information from different receptive fields, ultimately obtaining a multi-receptive field aggregated representation. The small receptive field describes rich details, corresponding to the features produced by the shallow layers of the network, and is represented by the output of the multi-resolution collaborative module. Characteristics of partial receptive fields Describes structural information between and within objects, through a The convolution is then performed through convolution in the channel direction. This achieves compression and dilation, aggregating information from neighboring channels. Convolution Structure such as Figure 4 As shown.

[0031] Deeper features The features are obtained through the same convolution, compression, and dilation processes. Features from different levels are aggregated together via residual connections to achieve multi-receptive-field feature aggregation. The computational cost is statistically analyzed using FLOPs (Floating-point Operations per Second), as given by the following formula: ; in, K h and K w These are the height and width of the convolution kernel, respectively. C in , C out These correspond to the number of input and output channels, respectively. H out , W out These correspond to the height and width of the output features, respectively.

[0032] In this embodiment, if C out Set to 64, H out , W out If both are set to 512, then the computational cost of this module can be calculated using the above formula to be 143.7 GFLOPs. This module... Convolution performs compression and dilation along the channel dimension, compared to directly using... The convolution method has a significant advantage in computational cost of 235.1 GFLOPs under the same conditions.

[0033] Specifically, the pixel attention weighting processing in S4 is implemented using a pixel attention mechanism in the reconstruction module, with the specific structure as follows: Figure 5 As shown.

[0034] This module primarily guides multi-resolution reconstruction using low-resolution, general information. Pixel attention extracts salient features from pixels, generating a three-dimensional weight matrix. As an attention feature, pixel attention generates attention coefficients for all pixels in the feature map. The reconstruction module based on the pixel attention mechanism consists of two... Attention weighting is achieved through convolution, LeakyReLU function, and Sigmoid function, and is implemented using a weight matrix. The correlations between pixels are fused. Then, a deconvolution layer is used to upsample the previous features with a set of deconvolution filters. Deconvolution can be seen as the inverse operation of convolution. For convolution, the filters are strided with the image. The convolution, where the output is the input. Similarly, if we swap the positions of the input and output, the output will be 1 / 2 the size of the input. The generalization of low resolution can guide the high-resolution reconstruction process. High-resolution features concatenate features of different resolutions individually, resulting in a... The convolutional layer maps channel information and compresses the number of channels. Finally, the output is summed element-wise with the input image at all resolutions to obtain the reconstructed image. image.

[0035] In this embodiment, the computational load of this module, calculated using the same conditions and formulas as described above, is 69.7 GFLOPs, compared to directly using... The convolution method has a significant advantage in computational cost of 114.8 GFLOPs under the same conditions.

[0036] Furthermore, this invention constructs an average L1 norm error based on the multi-resolution image reconstruction loss of the training set, and uses it as a loss function for training: ; In the formula, It is the first The number of pixels at each resolution. The loss is calculated as the average of the absolute differences between the model's predicted values ​​and the true target values. By simultaneously optimizing the outputs at three resolutions, the network is encouraged to learn a scale-invariant, sharp image representation. The final loss function... These are the three resolution outputs corresponding to each. The sum of losses.

[0037] The above is one embodiment of the UAV aerial image multi-resolution collaborative real-time deblurring method provided in this example. Based on the same idea, this example also provides a corresponding UAV aerial image multi-resolution collaborative real-time deblurring system: The data acquisition unit is used to acquire high-resolution, medium-resolution, and low-resolution images taken by the UAV. The multi-resolution collaborative guidance unit is used to extract shallow features from images of three resolutions; it is also used to upsample the shallow features of medium and low resolutions to the same size as the shallow features of high resolution, and then concatenate them to obtain high-resolution features; to downsample and upsample the shallow features of high resolution and low resolutions to the same size as the shallow features of medium resolution, and then concatenate them to obtain medium-resolution features; it is also used to upsample the medium-resolution features and concatenate them with the high-resolution features to obtain high-resolution concatenated features; to upsample the shallow features of low resolutions and concatenate them with the medium-resolution features to obtain medium-resolution concatenated features; and to sum the high-resolution concatenated features, medium-resolution features, and low-resolution shallow features element-wise with the original resolution image to obtain multi-resolution collaborative guidance features. The multi-receptive-field aggregation unit is used to sequentially extract features of different depths from multi-resolution collaboratively guided features, including convolution, compression and dilation, and aggregate the features of different depths through residual connections to obtain the multi-receptive-field aggregated representation. The deblurring image processing unit is used to perform pixel attention weighting on the multi-receptive field aggregation representation, and after upsampling the medium and low resolution weighted features, the three resolution features are concatenated and summed element-wise with the high resolution image to obtain the reconstructed high resolution deblurring image.

[0038] This embodiment also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The provided method is a multi-resolution collaborative real-time deblurring method for drone aerial images.

[0039] Those skilled in the art will understand that implementing all or part of the processes in the methods of the above embodiments can be accomplished by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0040] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for multi-resolution collaborative real-time deblurring of UAV aerial images, characterized in that, Includes the following steps: Acquire high-resolution, medium-resolution, and low-resolution images from drone aerial photography, and extract shallow features from the three resolution images. High-resolution features are obtained by upsampling medium-resolution and low-resolution shallow features to the same size as high-resolution shallow features and then concatenating them; medium-resolution features are obtained by downsampling and upsampling high-resolution and low-resolution shallow features to the same size as medium-resolution shallow features, respectively. High-resolution concatenated features are obtained by upsampling medium-resolution features and concatenating them with high-resolution features; medium-resolution concatenated features are obtained by upsampling low-resolution shallow features and concatenating them with medium-resolution features; multi-resolution collaborative guidance features are obtained by summing the high-resolution concatenated features, medium-resolution concatenated features, and low-resolution shallow features element-wise with the original resolution map. The multi-resolution collaborative guided features are sequentially extracted at different depths, including convolution, compression and dilation. The features at different depths are then aggregated through residual connections to obtain a multi-receptive field aggregated representation. The multi-receptive field aggregation representation is weighted by pixel attention, and after upsampling the weighted features at medium and low resolutions, the three resolution features are concatenated and summed element-wise with the high-resolution image to obtain the reconstructed high-resolution deblurred image. The multi-resolution collaborative guidance features are obtained through processing by a multi-resolution collaborative module, which includes three feature processing branches with different resolutions. The shallow feature extraction in the feature processing branch uses a shallow convolution module; In this process, the high-resolution and medium-resolution feature processing branches are convolutionally processed after cascading to obtain high-resolution and medium-resolution features; each feature processing branch undergoes a convolutional layer before element-wise summation with the original resolution image. The shallow convolution module employs... Convolution and Convolutional structure; high-resolution and medium-resolution feature processing branches are processed in convolution after cascading. and Convolution; The multi-receptive-field aggregation representation is implemented through a feature aggregation module; the feature aggregation module is implemented through... The convolution process is applied to the collaboratively guided features, through convolution in the channel direction. Compression and dilation are performed to obtain deep feature extraction; Multi-receptive field aggregation represents a reconstruction module that performs pixel attention weighting and employs a pixel attention mechanism. The reconstruction module of the pixel attention mechanism uses two Convolution, LeakyReLU function and sigmoid function are used to perform attention weighting, and the correlation between each pixel is fused through the weight matrix to obtain the weighted features; The medium and low resolution weighted feature upsampling uses a deconvolution layer, which uses a set of deconvolution filters to upsample the features.

2. The method for multi-resolution collaborative real-time deblurring of UAV aerial images according to claim 1, characterized in that, The medium- and low-resolution drone aerial images are obtained by sequentially convolving and downsampling the high-resolution images.

3. The method for multi-resolution collaborative real-time deblurring of UAV aerial images according to claim 1, characterized in that, The process involves weighting the multi-receptive-field aggregation representation with pixel attention, then summing the weighted features after upsampling at medium and low resolution with the original medium and low resolution images element-wise to obtain the reconstructed medium and low resolution deblurred images. Specifically, the average L1 norm error is constructed based on the multi-resolution image reconstruction loss of the training set and used as the loss function for training.

4. A deblurring system applying the multi-resolution collaborative real-time deblurring method for UAV aerial images as described in any one of claims 1-3, characterized in that, The system includes: The data acquisition unit is used to acquire high-resolution, medium-resolution, and low-resolution images taken by the drone. The multi-resolution collaborative guidance unit is used to extract shallow features from images of three resolutions; it is also used to upsample the shallow features of medium and low resolutions to the same size as the shallow features of high resolution, and then concatenate them to obtain high-resolution features; to downsample and upsample the shallow features of high resolution and low resolutions to the same size as the shallow features of medium resolution, and then concatenate them to obtain medium-resolution features; it is also used to upsample the medium-resolution features and concatenate them with the high-resolution features to obtain high-resolution concatenated features; to upsample the shallow features of low resolutions and concatenate them with the medium-resolution features to obtain medium-resolution concatenated features; and to sum the high-resolution concatenated features, medium-resolution concatenated features, and low-resolution shallow features element-wise with the original resolution image to obtain multi-resolution collaborative guidance features. The multi-receptive-field aggregation unit is used to sequentially extract features of different depths from multi-resolution collaboratively guided features, including convolution, compression and dilation, and aggregate the features of different depths through residual connections to obtain the multi-receptive-field aggregated representation. The deblurring image processing unit is used to perform pixel attention weighting on the multi-receptive field aggregation representation, and after upsampling the medium and low resolution weighted features, the three resolution features are concatenated and summed element-wise with the high resolution image to obtain the reconstructed high resolution deblurring image.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a data processing program, which, when executed by a processor, implements the multi-resolution collaborative real-time deblurring method for UAV aerial images as described in any one of claims 1 to 3.

6. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the multi-resolution collaborative real-time deblurring method for UAV aerial images as described in any one of claims 1 to 3.

Images