Deep learning super-resolution reconstruction system and training optimization method for unmanned aerial vehicle surveying and mapping image

Abstract

The invention relates to a deep learning super-resolution reconstruction system of an unmanned aerial vehicle surveying and mapping image and a training optimization method, particularly relates to the field of unmanned aerial vehicle surveying and mapping image processing, and remarkably improves the reconstruction precision of the unmanned aerial vehicle surveying and mapping image in a complex terrain scene by creatively fusing terrain elevation data and a deep learning technology. The problems of geometric distortion, elevation information splitting and error accumulation caused by topographic relief in the prior art are thoroughly solved. Through topographic adaptive acquisition and four-channel data fusion, the image reconstruction process can sense topographic relief changes; terrain projection distortion is dynamically compensated by using a deformable convolution mechanism, and a high-resolution image with accurate spatial position and rich texture details is reconstructed; and in combination with control point constraint training and a closed-loop quality inspection mechanism, the result is ensured to meet the national surveying and mapping precision standard, and the whole-process optimization upgrading from data acquisition to result output is realized.

Description

Technical Field

[0001] The present invention relates to the field of UAV surveying and mapping image processing, and more specifically, to a deep learning super-resolution reconstruction system and training optimization method for UAV surveying and mapping images. Background Art

[0002] In the field of drone mapping, low-altitude image acquisition often faces the problem of geometric distortion caused by complex terrain. Existing solutions are mainly divided into two categories: Geometric correction technology: uses affine transformation or collinear equations to perform planar correction on images, but cannot solve the nonlinear deformation caused by sudden changes in elevation.

[0003] Super-resolution reconstruction technology: Mainstream models such as EDSR improve resolution through residual learning, but the input data has been geometrically corrected by default and terrain information is not integrated.

[0004] Although this type of technology can improve resolution, the reconstructed images still have spatial coordinate deviations of >3 meters in areas with undulating terrain, such as mountainous areas and urban buildings.

[0005] When a drone collects images at an oblique angle in areas with sudden elevation changes (such as steep slopes and cliffs), the following problems may occur due to the coupling of flight attitude fluctuations and terrain undulations: Projection distortion: The same object on the surface appears offset in adjacent images, destroying the texture continuity of the super-resolution model; Fragmentation of elevation information: RGB images lack Z-axis coordinates, and the model cannot perceive the impact of terrain undulations on pixel displacement; Error accumulation: Geometric correction and super-resolution reconstruction are processed step by step. The linear assumption of affine transformation cannot correct nonlinear distortion (such as the displacement difference between the top and foundation of a building). Summary of the Invention

[0006] In response to the technical problems existing in the prior art, the present invention provides a deep learning super-resolution reconstruction system and training optimization method for UAV surveying and mapping images. By setting a data acquisition module, a terrain data fusion module, a terrain perception super-resolution module, an automatic quality inspection module and a feedback optimization module, the problems raised in the above background technology are solved.

[0007] The present invention solves the above technical problems with the following technical solutions: a deep learning super-resolution reconstruction system for UAV surveying and mapping images, specifically comprising: a data acquisition module, a terrain data fusion module, a terrain perception super-resolution module, an automatic quality inspection module, and a feedback optimization module; Data Acquisition Module: Dynamically plans flight parameters based on the terrain complexity index, simultaneously collects low-resolution images, lidar point clouds, and ground control point data, performs spatiotemporal registration on the collected multi-source data, generates a spatially aligned digital surface model and the original low-resolution UAV image dataset, and sends the dataset to the terrain data fusion module; Terrain data fusion module: normalizes the digital surface model in the acquired data set and generates a four-channel input tensor for splicing with the UAV image; Terrain-aware super-resolution module: This module uses a deformable convolutional layer to compensate for input data distortion and a residual dense network to perform super-resolution reconstruction to obtain high-resolution images. Automatic quality inspection module: calculates the information entropy improvement rate and control point residual of the weak texture area of ​​the reconstructed image and generates a quality assessment report; Feedback optimization module: Dynamically corrects acquisition parameters and network weights based on quality inspection results, and updates flight mission instructions; In a preferred embodiment, the calculation formula of the terrain complexity index is: ; in, represents the terrain complexity index, Indicates the total number of elevation points in the analysis area, Indicates the sum of all elevation points. Indicates the The elevation gradient of a point, Represents the standard deviation of the slope across the entire domain.

[0008] In a preferred embodiment, the specific steps of setting the flight parameters based on the calculated terrain complexity index value are: S1, if , then the flight altitude is 100m and the heading overlap rate is ≥80%; S2, if , then the flight altitude is 50m and the heading overlap rate is ≥90%; The data acquisition module also includes a spatial registration unit, a time synchronization unit and a control point projection unit, wherein: The spatial registration unit uses the ICP point cloud registration algorithm to align the lidar point cloud to the image coordinate system, with a plane error of ≤0.02m. The time synchronization unit uses the PTP1588 protocol to ensure that the timestamp deviation of multi-source data is ≤10ms. The control point projection unit corrects the coordinates of the ground control points based on the collinear equation, with an elevation error of ≤0.05m.

[0009] In a preferred embodiment, the specific formula for the normalization process is: ; in, Indicates the original elevation value, They represent the mean and standard deviation of the corresponding regional elevation respectively.

[0010] In a preferred embodiment, in the terrain-aware super-resolution module, the calculation formula for the three-dimensional offset of the deformable convolution layer is: ; in, Indicates the The three-dimensional offset of the sampling points, Express From 1 to , represents the learnable weights associated with each convolution kernel, represents the convolution kernel size, Represents the hyperbolic tangent function, which is used to perform nonlinear transformation on the calculation results. Represents a learnable weight matrix, whose dimension is 1×3, that is, 1 elevation input channel is mapped to 3 offset dimensions , and the output offset satisfies the constraint: , represents the convolution operation, Indicates the reference point, Represents the index of the convolution kernel The relevant data values, represents the bias term.

[0011] In a preferred embodiment, when the terrain-aware super-resolution module performs reconstruction, the loss function Constrained network training, the constraint expression is: ; in, The loss function of super-resolution, represents the geometric loss function, , represents the SIFT feature extraction function, represents the control point mask, represents the mask weight, constraining the error of high-importance regions, represents the loss function of the ground control points, , represents the image space projection function based on the collinearity equation, represents the true coordinates of the ground control points, Represents the weighting coefficient.

[0012] In a preferred embodiment, the calculation formula for the information entropy improvement rate of the weak texture area is: ; in, represents the entropy of the super-resolution image, represents the entropy of the low-resolution image, It means that the improvement rate of weak texture area is calculated by comparing the entropy difference between low-resolution image and super-resolution image; The calculation formula of the control point residual is: ; in, Represents the residual of the ground control point, that is, the error between the projection position of the super-resolution image and the actual position of the ground control point. represents the difference between the position of each control point in the super-resolution image and its true ground control point position, Represents a constant.

[0013] In a preferred embodiment, the strategy for generating a quality assessment report based on the calculated information entropy improvement rate of the weak texture area and the control point residual is: like , it means that the quality inspection has reached the standard and the output acceptance result is qualified. like , it means that the standard is not met and dynamic correction is required through the feedback optimization module, and an acceptance failure result is output; A surveying and mapping report in PDF format is automatically generated based on the acceptance results. The content of the surveying and mapping report includes a GCP residual distribution heat map; an MTF50 comparison curve before and after reconstruction of the mountainous area; and the blockchain summary value is written into the file metadata.

[0014] In a preferred embodiment, the dynamic correction strategy is: like , then increase the weight of weak texture samples; like , the flight overlap rate will be increased by 15%.

[0015] This application also provides a training optimization method for a deep learning super-resolution reconstruction system for UAV mapping images, which specifically includes the following steps: S101, constructing a training dataset including low-resolution images, digital surface models, high-resolution true values, and ground control point coordinates; S102, performing normalization processing on the digital surface model to generate a four-channel input tensor for image splicing; S103, calculating the three-dimensional offset through the deformable convolution layer, and generating a distortion compensation feature map according to the terrain data; S104, using a residual dense network to perform super-resolution reconstruction and output a high-resolution image; S105, calculating the information entropy improvement rate and the control point residual; S106. When the quality meets the standard, output the results; otherwise, return to S101 to adjust the acquisition parameters or update the network weights in S104.

[0016] The beneficial effects of the present invention are: by setting a data acquisition module, a terrain data fusion module, a terrain perception super-resolution module, an automatic quality inspection module and a feedback optimization module, the terrain elevation data and deep learning technology are innovatively integrated, the reconstruction accuracy of UAV surveying and mapping images in complex terrain scenes is significantly improved, and the geometric distortion, elevation information fragmentation and error accumulation problems caused by terrain undulations in the existing technology are completely solved: through terrain adaptive acquisition and four-channel data fusion, the image reconstruction process can perceive the changes in terrain undulations; the deformable convolution mechanism is used to dynamically compensate for terrain projection distortion, and reconstruct high-resolution images with precise spatial position and rich texture details; combined with control point constraint training and closed-loop quality inspection mechanism, it is ensured that the results meet the national surveying and mapping accuracy standards, and the whole process from data acquisition to results output is optimized and upgraded. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Flow chart of the method of the present invention; Figure 2 This is a system structure diagram of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0019] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the described features. In the description of this application, "plurality" means two or more, unless otherwise specifically specified.

[0020] In the description of this application, the term "for example" is used to mean "used as an example, illustration or explanation". Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is given to enable any person skilled in the art to implement and use the present invention. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art will recognize that the present invention can be implemented without using these specific details. In other examples, well-known structures and processes will not be elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest scope consistent with the principles and features disclosed in this application.

[0021] Example 1 This embodiment provides Figure 1 The present invention provides a training optimization method for a deep learning super-resolution reconstruction system for UAV mapping images, which specifically includes the following steps: S101. Determine flight altitude based on terrain complexity index, simultaneously collect images, lidar point clouds, and ground control points, and construct a training dataset consisting of low-resolution images, digital surface models, high-resolution ground truth, and ground control point coordinates. S102, performing normalization processing on the digital surface model to generate a four-channel input tensor for image splicing; S103, calculating the three-dimensional offset through the deformable convolution layer, and generating a distortion compensation feature map according to the terrain data; S104, using a residual dense network to perform super-resolution reconstruction and output a high-resolution image; S105, calculating the information entropy improvement rate and the control point residual; S106. When the quality meets the standard, output the results; otherwise, return to S101 to adjust the acquisition parameters or update the network weights in S104.

[0022] Example 2 This embodiment provides Figure 2 The deep learning super-resolution reconstruction system for UAV mapping images shown in the figure specifically includes: a data acquisition module, a terrain data fusion module, a terrain perception super-resolution module, an automatic quality inspection module and a feedback optimization module; Data Acquisition Module: Dynamically plans flight parameters based on the terrain complexity index, simultaneously collects low-resolution images, lidar point clouds, and ground control point data, performs spatiotemporal registration on the collected multi-source data, generates a spatially aligned digital surface model and the original low-resolution UAV image dataset, and sends the dataset to the terrain data fusion module; Terrain data fusion module: normalizes the acquired data into a digital surface model and generates a four-channel input tensor for splicing with drone images. , in this formula: Represents the original low-resolution drone image, shaped like , Represents the normalized digital surface model, the shape is , Represents stitching low-resolution imagery and elevation data, resulting in a shape of The four-channel input tensor of Represents the final four-channel input tensor with shape ), as the network input, the purpose of this module design is to use elevation information as a priori input of spatial position, breaking the traditional RGB three-channel limitation; Terrain-aware super-resolution module: This module uses a deformable convolution layer at the encoder front end to compensate for input data distortion, allowing the convolution kernel to adaptively adjust feature sampling positions based on local terrain undulations (for example, increasing vertical sampling spacing in steep slopes). This module then uses a residual dense network for super-resolution reconstruction, resulting in high-resolution images. Automatic quality inspection module: calculates the information entropy improvement rate and control point residual of the weak texture area of ​​the reconstructed image and generates a quality assessment report; Feedback optimization module: Dynamically corrects acquisition parameters and network weights based on quality inspection results, and updates flight mission instructions.

[0023] In this embodiment, the data acquisition module needs to be specifically explained. The calculation formula of the terrain complexity index is: ; in, Represents the terrain complexity index. The larger the value, the more complex the terrain. It is the core indicator for flight altitude decision-making (>0.3 indicates high-risk terrain). Indicates the total number of elevation points in the analysis area, which is determined by the DEM resolution (e.g. 1km at 1m resolution). 2 Region N = 1,000,000), Indicates the summation of all elevation points, that is, the cumulative calculation of all grid points. Indicates the The elevation gradient of a point is calculated as follows: , which represents a function exist The magnitude of the gradient in the direction, It represents the absolute value of the elevation gradient, that is, it removes the directional influence and quantifies the steepness of the slope (>30° is a steep slope). It represents the standard deviation of the global slope, and its calculation formula is: , Indicates the The slope value of the point, , represents the average slope of the entire domain, ; The specific steps for setting flight parameters based on the calculated terrain complexity index value are as follows: S1, if , which means flat / gentle slope terrain, the flight altitude is 100m and the heading overlap rate is ≥80%; S2, if , which refers to steep slope / canyon terrain, the flight altitude is 50m and the heading overlap rate is ≥90%; The data acquisition module also includes a spatial registration unit, a time synchronization unit, and a control point projection unit, wherein: The spatial registration unit uses the ICP point cloud registration algorithm to align the lidar point cloud to the image coordinate system, with a plane error of ≤0.02m. The time synchronization unit uses the PTP1588 protocol to ensure that the timestamp deviation of multi-source data is ≤10ms. The control point projection unit corrects the coordinates of the ground control points based on the collinear equation, with an elevation error of ≤0.05m.

[0024] In this embodiment, it is necessary to explain in detail the terrain data fusion module, and the specific formula for normalization processing is: ; in, Indicates the original elevation value, They represent the mean and standard deviation of the corresponding area elevation, respectively, and are used to map the elevation value to the [-1,1] interval.

[0025] In this embodiment, it is specifically necessary to explain that in the terrain-aware super-resolution module, the calculation formula for the three-dimensional offset of the deformable convolution layer is: ; in, Indicates the The three-dimensional offset of the sampling point (including the offset and scaling factor in the x and y directions, that is, the offset learned by the model through convolution operations), Express From 1 to , Represents the learnable weights associated with each convolution kernel, which is used to adjust the contribution of each convolution kernel to the offset. represents the convolution kernel size, Represents the hyperbolic tangent function, which is used to perform nonlinear transformation on the calculation results. Its output range is [-1,1], which is used to limit the size of the offset to avoid excessive distortion. Represents a learnable weight matrix, whose dimension is 1×3, that is, 1 elevation input channel is mapped to 3 offset dimensions , and the output offset satisfies the constraint: , represents the convolution operation, Indicates a reference point or initial point, which is a starting position of the input image or data, used to calculate the offset. Represents the index of the convolution kernel The associated data value, usually the value of a pixel or position in an image or feature map, Represents the bias term, which is usually the same as the bias in a neural network, and is used as an additive term to adjust the output of the model; When the terrain-aware super-resolution module performs reconstruction, the loss function Constrained network training, the constraint expression is: ; in, The super-resolution loss function is usually used to measure the difference between the low-resolution image output by the network and the high-resolution target image. Represents the geometric loss function, that is, the difference between the SIFT features of the super-resolution image and the high-resolution image. By minimizing this loss function, the model can optimize the reconstruction of geometric details. , represents the SIFT feature extraction function, represents the control point mask (marking the key spatial locations of ridge lines and building corners), Represents mask weighting, constraining the error of high-importance regions. This formula is designed to force the spatial coordinates of the reconstructed image at the control point to align with the high-resolution true value, with a deviation of ≤0.5 pixels. The loss function for the ground control points is the difference between the projected position in the super-resolution image and the true position of the ground control points based on the image space projection method. This loss function helps ensure that the reconstructed terrain matches the actual ground control points. , represents the image space projection function based on the collinearity equation, represents the true coordinates of the ground control points, Represents the weighting coefficient, which is used to adjust the importance of different loss functions. In addition, during reconstruction, the terrain-aware super-resolution module uses a 16-layer residual dense network to upsample the distortion compensation features, and the output resolution is increased to a high-resolution image with a resolution twice that of the original image.

[0026] In this embodiment, the automatic quality inspection module specifically needs to be explained. The calculation formula for the information entropy improvement rate of the weak texture area is: ; in, represents the entropy (information entropy) of the super-resolution image, represents the entropy (information entropy) of the low-resolution image, It means that the improvement rate of weak texture areas is calculated by comparing the entropy difference between low-resolution images and super-resolution images. Information entropy is an indicator to measure the complexity of information in an image. A higher entropy value means that the image contains more details. The calculation formula for the control point residual is: ; in, Represents the residual of the ground control point, that is, the error between the projection position of the super-resolution image and the actual position of the ground control point. represents the difference between the position of each control point in the super-resolution image and its true ground control point position, Indicates the size of this difference, that is, the position error, Represents a constant by the reciprocal To adjust the residuals, usually to normalize the errors to match the scale of the actual control points; Based on the calculated information entropy improvement rate of the weak texture area and the control point residual, the strategy for generating the quality assessment report is: like , it means that the quality inspection has reached the standard and the output acceptance result is qualified. like , it means that the standard is not met and dynamic correction is required through the feedback optimization module, and an acceptance failure result is output; A surveying and mapping report in PDF format is automatically generated based on the acceptance results. The content of the surveying and mapping report includes a GCP residual distribution heat map; an MTF50 comparison curve before and after reconstruction of the mountainous area; and the blockchain summary value (SHA-256) is written into the file metadata.

[0027] In this embodiment, it is specifically necessary to explain the feedback optimization module, and the dynamic correction strategy is: like , then increase the weight of weak texture samples; like , the flight overlap rate will be increased by 15%.

[0028] It should be noted that, in the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0029] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0030] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts 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, a special-purpose computer, an embedded computer, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0031] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0032] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0033] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0034] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A deep learning super-resolution reconstruction system for UAV mapping images, characterized by: Specifically include: Data acquisition module, terrain data fusion module, terrain perception super-resolution module, automatic quality inspection module and feedback optimization module; Data Acquisition Module: Dynamically plans flight parameters based on the terrain complexity index, simultaneously collects low-resolution images, lidar point clouds, and ground control point data, performs spatiotemporal registration on the collected multi-source data, generates a spatially aligned digital surface model and the original low-resolution UAV image dataset, and sends the dataset to the terrain data fusion module; Terrain data fusion module: normalizes the digital surface model in the acquired data set and generates a four-channel input tensor for splicing with the UAV image; Terrain-aware super-resolution module: This module uses a deformable convolutional layer to compensate for input data distortion and a residual dense network to perform super-resolution reconstruction to obtain high-resolution images. Automatic quality inspection module: calculates the information entropy improvement rate and control point residual of the weak texture area of ​​the reconstructed image and generates a quality assessment report; Feedback optimization module: Dynamically corrects acquisition parameters and network weights based on quality inspection results, and updates flight mission instructions.

2. The deep learning super-resolution reconstruction system for UAV surveying and mapping images according to claim 1 is characterized by: The calculation formula of the terrain complexity index is: ； in, represents the terrain complexity index, Indicates the total number of elevation points in the analysis area, Indicates the sum of all elevation points. Indicates the The elevation gradient of a point, Represents the standard deviation of the slope across the entire domain.

3. The deep learning super-resolution reconstruction system for UAV surveying and mapping images according to claim 2, characterized in that: The specific steps for setting flight parameters based on the calculated terrain complexity index value are as follows: S1, if , then the flight altitude is 100m and the heading overlap rate is ≥80%; S2, if , then the flight altitude is 50m and the heading overlap rate is ≥90%; The data acquisition module also includes a spatial registration unit, a time synchronization unit and a control point projection unit, wherein: The spatial registration unit uses the ICP point cloud registration algorithm to align the lidar point cloud to the image coordinate system, with a plane error of ≤0.02m. The time synchronization unit uses the PTP1588 protocol to ensure that the timestamp deviation of multi-source data is ≤10ms. The control point projection unit corrects the coordinates of the ground control points based on the collinear equation, with an elevation error of ≤0.05m.

4. The deep learning super-resolution reconstruction system for UAV surveying and mapping images according to claim 3 is characterized by: The specific formula for the normalization process is: ; in, Indicates the original elevation value, They represent the mean and standard deviation of the corresponding regional elevation respectively.

5. The deep learning super-resolution reconstruction system for UAV surveying and mapping images according to claim 4 is characterized by: In the terrain-aware super-resolution module, the calculation formula for the three-dimensional offset of the deformable convolution layer is: ； in, Indicates the The three-dimensional offset of the sampling points, Express From 1 to , represents the learnable weights associated with each convolution kernel, represents the convolution kernel size, Represents the hyperbolic tangent function, which is used to perform nonlinear transformation on the calculation results. Represents a learnable weight matrix, whose dimension is 1×3, that is, 1 elevation input channel is mapped to 3 offset dimensions , and the output offset satisfies the constraint: , represents the convolution operation, Indicates the reference point, Represents the index of the convolution kernel The relevant data values, represents the bias term.

6. The deep learning super-resolution reconstruction system for UAV surveying and mapping images according to claim 5, characterized in that: When the terrain-aware super-resolution module performs reconstruction, the loss function Constrained network training, the constraint expression is: ； in, The loss function of super-resolution, represents the geometric loss function, , represents the SIFT feature extraction function, represents the control point mask, represents the mask weight, constraining the error of high-importance regions, represents the loss function of the ground control points, , represents the image space projection function based on the collinearity equation, represents the true coordinates of the ground control points, Represents the weighting coefficient.

7. The deep learning super-resolution reconstruction system for UAV surveying and mapping images according to claim 6, characterized in that: The calculation formula for the information entropy improvement rate of the weak texture area is: ； in, represents the entropy of the super-resolution image, represents the entropy of the low-resolution image, It means that the improvement rate of weak texture area is calculated by comparing the entropy difference between low-resolution image and super-resolution image; The calculation formula of the control point residual is: ； in, Represents the residual of the ground control point, that is, the error between the projection position of the super-resolution image and the actual position of the ground control point. represents the difference between the position of each control point in the super-resolution image and its true ground control point position, Represents a constant.

8. The deep learning super-resolution reconstruction system for UAV surveying and mapping images according to claim 7, characterized in that: Based on the calculated information entropy improvement rate of the weak texture area and the control point residual, the strategy for generating the quality assessment report is: like , it means that the quality inspection has reached the standard and the output acceptance result is qualified. like , it means that the standard is not met and dynamic correction is required through the feedback optimization module, and an acceptance failure result is output; A surveying and mapping report in PDF format is automatically generated based on the acceptance results. The content of the surveying and mapping report includes a GCP residual distribution heat map; an MTF50 comparison curve before and after reconstruction of the mountainous area; and the blockchain summary value is written into the file metadata.

9. The deep learning super-resolution reconstruction system for UAV surveying and mapping images according to claim 8, characterized in that: The dynamic correction strategy is: like , then increase the weight of weak texture samples; like , the flight overlap rate will be increased by 15%.

10. A training optimization method for a deep learning super-resolution reconstruction system for UAV surveying and mapping images based on any one of claims 1 to 9, characterized in that: The following steps are involved: S101, constructing a training dataset including low-resolution images, digital surface models, high-resolution true values, and ground control point coordinates; S102, performing normalization processing on the digital surface model to generate a four-channel input tensor for image splicing; S103, calculating the three-dimensional offset through the deformable convolution layer, and generating a distortion compensation feature map according to the terrain data; S104, using a residual dense network to perform super-resolution reconstruction and output a high-resolution image; S105, calculating the information entropy improvement rate and the control point residual; S106. When the quality meets the standard, output the results; otherwise, return to S101 to adjust the acquisition parameters or update the network weights in S104.

Images