Aerial survey image deblurring processing method based on multi-frame fusion algorithm

Abstract

This invention relates to the field of image processing technology, specifically to a method for deblurring aerial survey images based on a multi-frame fusion algorithm. First, a pixel-level motion trajectory model of the image plane is constructed, trajectory parameters are extracted to generate a spatially varying blur kernel, and a global blur intensity factor is calculated. Then, candidate frames are extracted based on the ground projection overlap rate, and the optimal fusion frame subset is obtained through multi-dimensional quality quantization scoring and adaptive threshold selection. Subsequently, light and heavy blur regions are divided using a global average blur displacement threshold. Light blur regions are fused using rigid alignment and quality-weighted fusion, while heavy blur regions are fused using pixel-level offset alignment and mean fusion. Finally, image optimization is completed through adaptive noise reduction and dynamic detail enhancement, accurately adapting to the non-uniform motion blur of aerial survey images. The scientific candidate frame selection and high-precision partition alignment fusion effectively preserve image details, suppress artifacts and noise, and significantly improve the clarity and fidelity of aerial survey images, meeting the requirements of high-precision surveying.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to a method for deblurring aerial survey images based on a multi-frame fusion algorithm. Background Technology

[0002] During drone and airborne aerial surveying operations, cameras are prone to attitude and position changes during exposure, resulting in non-uniform motion blur in the images. Traditional deblurring methods often use a globally uniform blur kernel, which cannot adapt to pixel-level blur differences. Furthermore, when fusing multiple frames, the filtering, alignment, and weighting strategies are simplistic, easily leading to problems such as loss of detail, artifacts, and noise amplification, making it difficult to meet the requirements for sharpening high-precision aerial surveying images.

[0003] Traditional image deblurring methods often process single-frame images and use a globally uniform blur kernel for deconvolution operations. This approach cannot adapt to pixel-level blur differences in aerial survey images and is prone to problems such as edge artifacts, texture distortion, and noise amplification. Existing multi-frame fusion deblurring techniques rely solely on the ground overlap rate of images to select candidate frames without combining blur degree and image quality for quantitative scoring. Furthermore, they often employ a globally rigid alignment strategy, which suffers from insufficient alignment accuracy in heavily blurred areas. The fusion weights are also mostly fixed and cannot be adaptively allocated based on image quality.

[0004] In addition, existing technologies generally use globally uniform noise reduction and detail enhancement parameters in the post-processing stage of fusion. Lightly blurred areas are prone to losing real details due to excessive noise reduction, while heavily blurred areas will produce artifacts and noise amplification due to inappropriate enhancement intensity. Ultimately, the deblurred image is difficult to balance detail preservation, noise suppression, and edge realism, and cannot meet the processing requirements of high-precision aerial surveying and mapping for image clarity and fidelity. Summary of the Invention

[0005] To address the aforementioned shortcomings of existing technologies, this invention provides a method for deblurring aerial survey images based on a multi-frame fusion algorithm, which can effectively solve the problems of difficulty in repairing non-uniform blur in aerial survey images and poor fusion processing results in the background technology.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a method for deblurring aerial survey images based on a multi-frame fusion algorithm, comprising:

[0007] The system acquires the position and attitude information of the aerial survey platform and the aerial survey image sequence, performs data preprocessing and spatiotemporal alignment, and obtains the continuous camera motion trajectory function, the preprocessed image sequence, and the camera intrinsic parameters.

[0008] The location information includes longitude, latitude, and altitude, and the attitude information includes pitch angle, roll angle, and heading angle.

[0009] Position and attitude information within the exposure time interval of a single frame of aerial survey imagery is extracted, a pixel-level motion trajectory model of the image plane is constructed, a spatial variation blur kernel of the entire aerial survey imagery is generated by parsing, and the global blur intensity factor is calculated.

[0010] Candidate frame sets are extracted based on the overlap of image ground projection. The candidate frames are then subjected to multi-dimensional quality quantification scoring. The optimal fused frame subset is selected by adaptive quality threshold discrimination and optimal frame quantity constraint.

[0011] The global average blur displacement threshold is calculated based on the total displacement amplitude of the pixel-by-pixel trajectory of the current frame to be processed. Alignment fusion is then performed on the optimal fusion frame subset according to the light and heavy blur distinctions to obtain the initial fused deblurred image.

[0012] Based on image noise characteristics and global blur intensity factor, the initial fused image is optimized for noise reduction and detail enhancement, and finally the cleared aerial survey image is output.

[0013] Preferably, the continuous camera motion trajectory function, the preprocessed image sequence, and camera intra-camera parameters are obtained through data preprocessing and spatiotemporal alignment, including:

[0014] Outlier removal and Kalman filtering smoothing are performed on the position and attitude data collected by the aerial survey platform's positioning and attitude determination system. The continuous camera motion trajectory function is then obtained by cubic spline interpolation fitting.

[0015] Radiometric correction and nonlocal mean denoising were performed sequentially on the original aerial survey images to extract the camera intra-frame parameters and exposure time intervals of each image.

[0016] Ground control points within the aerial survey area are selected, and the time offset is optimized with the goal of minimizing the sum of squared residuals between the projected coordinates of the control points and the actual observed coordinates. The timestamps of the motion data are unified to the shutter time reference of the aerial camera to complete the spatiotemporal alignment.

[0017] Preferably, constructing a pixel-level motion trajectory model within the image plane specifically includes:

[0018] The exposure time range for each frame of the image is defined based on the start and end times of the exposure.

[0019] By calculating instantaneous position and attitude information, the instantaneous position and attitude information of the aerial camera at each moment within the exposure time interval are obtained.

[0020] Based on the instantaneous longitude, instantaneous latitude, and instantaneous altitude of the aerial camera's position, the latitude and longitude coordinates are converted into Cartesian coordinates to obtain the instantaneous coordinates of the aerial camera's center in the Cartesian coordinate system.

[0021] By combining the airflow disturbance compensation coefficient and the instantaneous height, the corrected instantaneous height, i.e., the instantaneous position information, is calculated.

[0022] The instantaneous pitch angle, instantaneous roll angle, and instantaneous yaw angle of the camera attitude are obtained through cubic spline interpolation.

[0023] By combining the preset sampling frequency and exposure time interval, a sliding window is calculated. Combined with the attitude coupling correction coefficient and the preset time interval, the corrected instantaneous attitude information is obtained.

[0024] Based on the corrected instantaneous position and attitude angle of the camera, combined with the intra-camera parameters, including focal length, pixel size and principal point offset, the instantaneous motion trajectory of each pixel in the image plane of each frame is constructed.

[0025] Integrating over all instants during the exposure period yields the complete motion trajectory of a single pixel throughout the entire exposure. By traversing all pixels in the corresponding plane of each frame, a pixel-level motion trajectory model for each frame is obtained.

[0026] Preferably, the process of parsing and generating a spatially varying fuzzy kernel and calculating the global fuzzy intensity factor includes:

[0027] For each pixel in a single frame image, the total displacement amplitude of the trajectory is extracted by synthesizing and calculating the total offset of the horizontal and vertical coordinates of that pixel during the exposure period.

[0028] Based on the directional relationship between the horizontal and vertical instantaneous motion speeds of the pixel during the exposure period, the main trajectory direction angle is extracted.

[0029] The trajectory discrete factor is extracted by statistically calculating the displacement increment distribution of the trajectory time sampling interval within the exposure period.

[0030] The size of the fuzzy kernel is determined by symmetric expansion and rounding operations based on the total displacement amplitude of the trajectory. The main direction angle of the trajectory is used as the main axis of fuzzy extension, and the coordinates within the kernel are rotated and mapped. The weight distribution within the kernel is calculated by combining the trajectory discrete factor and the weight normalization is completed. A spatial variation fuzzy kernel that fits the motion fuzzy characteristics of aerial surveying is generated with the current pixel as the center.

[0031] The total displacement amplitude of the trajectory of all pixels in the whole frame is statistically analyzed and normalized to obtain the global blur intensity factor of the single frame image.

[0032] Preferably, the candidate frame set is extracted based on image overlap, including:

[0033] Obtain the center height of the aerial camera, the focal length, image size, and pixel size of the camera's intrinsic parameters corresponding to a single frame of image, and calculate the total ground projection area corresponding to each frame of image.

[0034] The four corner points of the image of the frame to be processed and the adjacent frame are projected onto the same ground rectangular coordinate system to form a ground projection polygon. The intersection area of ​​the two is calculated. The ground projection overlap rate of the two images is obtained by combining the total ground projection area and the intersection area.

[0035] An adaptive overlap threshold is constructed by combining the global blur intensity factor of a single frame image, and adjacent frames that meet the overlap threshold requirements are selected to form an initial candidate frame set.

[0036] Preferably, a multi-dimensional quality quantification score for candidate frames is performed by combining fuzzy priors, including:

[0037] The inputs are the candidate frame set, the global blur intensity factor of each frame image, the total displacement amplitude of the pixel-by-pixel trajectory, and the noise residual of the preprocessed image.

[0038] The average blur displacement amplitude of the candidate frame is calculated by combining the image size and the total displacement amplitude of the pixel-by-pixel trajectory. Based on this average blur displacement amplitude, the blur fit score of a single frame image is calculated.

[0039] By combining the spatial gradient entropy, the frequency high-frequency energy entropy, and the fixed balance coefficient, the joint spatial-frequency sharpness entropy of a single frame image is obtained.

[0040] The noise suppression score of a single frame image is obtained by combining the average noise residual of the candidate frame set with the maximum average noise residual.

[0041] The fuzziness fit score, spatial-frequency joint sharpness entropy, and noise suppression score are weighted and summed using preset weighting factors to obtain the comprehensive quality score of a single frame image in the candidate frame set.

[0042] Preferably, the quality threshold discrimination and optimal fusion frame subset determination include:

[0043] By combining the clear image baseline quality threshold, the blur adaptive adjustment coefficient, and the global blur intensity factor of the frame to be processed, a comprehensive quality score threshold that adjusts with the degree of blur is constructed.

[0044] Set a constraint range for the optimal number of fused frames, sort the candidate frames according to the comprehensive quality score, and select candidate frames with a comprehensive quality score ≥ the comprehensive quality score threshold.

[0045] If the number of qualified candidate frames is lower than the lower limit of the constraint interval, the overall quality score is lowered by a fixed step size and the number of frames is replenished by re-screening.

[0046] If the number of candidate frames that meet the conditions is higher than the upper limit of the constraint interval, then the upper limit of the constraint interval is used as the selection criterion, and the frame image with the highest comprehensive quality score is selected to obtain the optimal fused frame subset.

[0047] Preferably, calculating the global average blur displacement threshold and performing partition alignment fusion on the optimal fusion frame subset includes:

[0048] The global average blur displacement threshold of the current frame to be processed is obtained by performing a global statistical average calculation based on the total displacement amplitude of the trajectory of all pixels in the frame to be processed.

[0049] Using the global average blur displacement threshold as the region discrimination criterion, the entire image domain of the frame to be processed is divided into lightly blurred regions and heavily blurred regions.

[0050] For lightly blurred areas, the fusion weight of each fused frame is calculated by combining the comprehensive quality score of the corresponding single frame image in the candidate frame set. Then, the pixel gray value of each pixel after fusion is calculated by combining the gray value of each pixel corresponding to the fused image of each frame.

[0051] For heavily blurred regions, the corresponding pixel displacement is calculated based on the total displacement amplitude of the pixel trajectory. Then, the corresponding pixels in each candidate frame are offset and aligned according to the pixel displacement. Finally, the arithmetic mean of the aligned pixel values ​​of multiple frames is calculated to obtain the fused pixel value of the heavily blurred region.

[0052] For the optimal fused frame subset after partition alignment, corresponding weighted fusion processing is performed on the lightly blurred and heavily blurred regions respectively, and the initial fused deblurred image is finally obtained.

[0053] Preferably, adaptive noise reduction and dynamic detail enhancement optimization are performed on the initial fused image, including:

[0054] The system takes the initial fused deblurred image, the global blur intensity factor of the frame to be processed, the preprocessed image noise residual, and the preset basic enhancement coefficient as inputs, and first performs adaptive noise reduction processing.

[0055] By combining the noise residual of the initial fused image with the global blur intensity factor of the frame to be processed, an adaptive noise reduction intensity coefficient that dynamically adjusts with the degree of blur is calculated, and the initial fused image is then subjected to regional noise reduction processing based on this coefficient.

[0056] Then, dynamic detail enhancement processing is performed. The dynamic enhancement coefficient is calculated by combining the global blur intensity factor of the frame to be processed with the preset basic enhancement coefficient. Based on this coefficient, the details of the denoised fused image are enhanced.

[0057] After noise reduction and detail enhancement optimization, the aerial survey images are output with improved clarity.

[0058] The technical solution provided by this invention has the following advantages compared with the known prior art:

[0059] 1. In the process of constructing the pixel-level motion trajectory model, the present invention performs projection transformation and height correction on the instantaneous position of the camera, performs sliding window mean and coupling correction on the attitude angle, and calculates the instantaneous pixel offset by integrating the camera intrinsic parameters. This is beneficial for accurately restoring the real motion trajectory of pixels during exposure and accurately characterizing the distribution characteristics of non-uniform blur in aerial survey images.

[0060] 2. In the process of generating spatially varying fuzzy kernels, this invention extracts trajectory parameters based on the total offset of pixel horizontal and vertical coordinates, the direction of motion velocity, and the distribution of displacement increments. Then, based on the trajectory features, it determines the size, principal axis, and weight of the fuzzy kernel. This is beneficial for abandoning the traditional globally unified fuzzy kernel and realizing pixel-by-pixel adaptive fuzzy modeling.

[0061] 3. In the candidate frame extraction process, the ground projection overlap rate is calculated by using camera intrinsic parameters and flight altitude, and an adaptive overlap threshold is constructed by combining the fuzzy intensity factor. This helps to eliminate adjacent frames with insufficient overlap and poor matching, thus ensuring the effective fusion value of candidate frames.

[0062] 4. In the multi-dimensional quality scoring process, the embodiments of the present invention use the weighted calculation of fuzzy adaptation degree, spatial-frequency joint sharpness entropy and noise suppression score to objectively quantify the quality of a single frame image and provide a scientific basis for the allocation of fusion weights. In the process of selecting the optimal fusion frame, the adaptive quality threshold and quantity constraint range are dynamically adjusted to ensure a stable number of fusion frames and optimal quality, and to avoid insufficient or redundant frames.

[0063] 5. In the partition alignment fusion process of this invention, light and heavy blur regions are divided by a blur displacement threshold, and rigid alignment weighted fusion and pixel-level offset alignment mean fusion are adopted respectively. This is beneficial to both preserve details in light blur regions and eliminate ghosting in heavy blur regions, thereby improving fusion accuracy. In the post-image optimization process, noise reduction and enhancement coefficients are dynamically calculated by blur intensity factor, which is beneficial to achieve regional adaptive processing, avoid loss of details, artifacts and noise amplification, and significantly improve the clarity and fidelity of aerial survey images. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0065] Figure 1 This is a schematic diagram of the implementation steps of the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0067] The present invention will be further described below with reference to embodiments.

[0068] Please see Figure 1 As shown, a method for deblurring aerial survey images based on a multi-frame fusion algorithm includes at least the following:

[0069] S1. Acquire the position information, attitude information and aerial survey image sequence of the aerial survey platform during image exposure, and perform preprocessing and spatiotemporal alignment.

[0070] In one specific embodiment, the motion data of the aerial survey aircraft and the image sequence captured by the aerial survey camera are preprocessed and spatiotemporally aligned, specifically as follows:

[0071] The positioning and attitude determination system onboard the aerial survey aircraft records motion data during flight at a preset sampling frequency. The motion data includes position information, latitude, and altitude, and attitude information, including pitch angle, roll angle, and heading angle. Each data point is accompanied by a precise timestamp. First, outliers in the raw motion data are detected and removed. Kalman filtering is used to smooth the data after outlier removal to eliminate high-frequency noise. Then, cubic spline interpolation is performed on the six components of the smoothed motion data (longitude, latitude, altitude, pitch angle, roll angle, and heading angle) to fit the discretely sampled motion data into a continuous camera motion trajectory function.

[0072] During flight, the aerial survey camera takes pictures of the ground at preset time intervals, with each exposure lasting a short time window, obtaining a sequence of raw aerial survey images. Radiometric correction is performed on the raw images using a radiometric response curve based on laboratory calibration to eliminate uneven image brightness caused by illumination variations and lens vignetting. Then, a nonlocal mean denoising algorithm is used to denoise the images, removing sensor noise while preserving image texture details. Finally, camera intrinsic parameters, including focal length, pixel size, and principal point offset, are extracted from the metadata of each frame, along with the exposure start and end times of that frame, which serve as the time basis for determining the motion data sampling interval in subsequent steps.

[0073] After preprocessing, the preprocessed image sequence and the camera parameters and exposure time range corresponding to each frame are output.

[0074] Ground control points with distinctive features in the aerial survey area are selected, and these control points can be detected in multiple consecutive image frames. For each image frame, the ground control points are projected onto the image plane using known exterior orientation elements to obtain the coordinates of the projected points. At the same time, the actual observed coordinates of the control points are extracted from the image using a feature matching algorithm. The time offset is optimized by minimizing the sum of squared residuals between the projected coordinates and observed coordinates of all control points. After calibration, the timestamps of all motion data are uniformly converted to the shutter time reference of the aerial camera.

[0075] This completes preprocessing and spatiotemporal alignment, thereby obtaining the continuous motion trajectory function corresponding to the image exposure time, the preprocessed image sequence, and the aerial camera parameters of each frame.

[0076] It should be noted that the preset sampling frequency is set according to the accuracy requirements of the aerial survey mission and the motion characteristics of the aerial survey platform. During flight, the aerial survey aircraft will be affected by factors such as airflow disturbances, resulting in high-frequency jitter. In order to accurately capture the minute movements of the camera during the exposure, the sampling frequency of the positioning and attitude determination system should satisfy the Nyquist sampling theorem, that is, the sampling frequency is at least twice the highest frequency component of the camera motion. In practical applications, the sampling frequency of the positioning and attitude determination system is usually set to 100Hz to 200Hz. For example, in this embodiment, it is set to 200Hz, that is, 200 sets of motion data are recorded per second, corresponding to a sampling interval of 5ms.

[0077] The preset time interval is set according to the ground resolution and forward overlap requirements of the aerial survey mission. Specifically, it is calculated by the aerial survey design software based on the aerial photography scale, ground coverage width, and the set forward overlap (usually 60% to 80%). For example, in this embodiment, the aerial survey camera takes continuous pictures at a time interval of 1 second to ensure that there is sufficient overlap between adjacent images.

[0078] The specific fitting process for fitting discrete sampled motion data into a continuous camera motion trajectory function is as follows: the six components of the smoothed motion data are interpolated using the cubic spline interpolation method, so that the fitted piecewise polynomial passes through the original data value at each sampling point and has continuous first and second derivatives at internal nodes, thereby obtaining a continuous camera motion trajectory function that can be directly calculated at any time.

[0079] The distinctive ground control points refer to ground features within the aerial survey area that have clear, stable, and easily identifiable features, such as road intersections, building corners, and field ridge intersections. These ground features have high contrast and local uniqueness in aerial survey images.

[0080] The known exterior orientation elements refer to six parameters that describe the spatial position and attitude of the aerial camera relative to the ground coordinate system at the moment of exposure. These include three linear elements (longitude, latitude, and altitude of the camera center) and three angular elements (pitch angle, roll angle, and heading angle). The exterior orientation elements are directly provided by the positioning and attitude determination system.

[0081] The feature matching algorithm refers to an algorithm that extracts local feature descriptors from images and performs cross-image matching, such as scale-invariant feature transformation algorithm or accelerated robust feature algorithm. By calculating the similarity between feature descriptors, the image plane coordinates of the same ground control point are located in different frames of images.

[0082] The positioning and attitude determination system is a high-precision navigation system composed of a global satellite navigation system and an inertial measurement unit through a data fusion processing unit. In this scheme, it is used to record the position and attitude information of the aerial survey aircraft during flight at a preset sampling frequency, providing raw data input for subsequent pixel-level fuzziness quantification based on motion data.

[0083] S2. Extract camera motion data within the exposure time interval of a single frame of aerial survey image, construct a pixel-level motion trajectory model in the image plane, and generate a spatial variation blur kernel for the entire aerial survey image based on the pixel-level motion trajectory analysis.

[0084] In one specific embodiment, the process of constructing a pixel-level motion trajectory model in the image plane is as follows: Based on the exposure start and end times of a preprocessed single-frame aerial survey image, the exposure time interval of that frame image is defined as... ,in The starting point of the exposure. The end time of the exposure. and All of these are exposure time information extracted from the corresponding image metadata in step S1, including exposure duration. .

[0085] Based on the motion trajectory function of the continuous aerial camera in step S1, extract... The camera motion data within the time interval yields the instantaneous motion sequence during the exposure period, including position information and attitude information. The position information is the instantaneous coordinates of the camera center in the ground coordinate system, and the attitude information is the instantaneous attitude angle of the camera.

[0086] Define the instantaneous time variable within the exposure period as , Based on the cubic spline interpolation function in step S1, the solution is quantized. The instantaneous motion parameters of the aerial camera at any given moment are calculated, specifically including instantaneous position information calculation and instantaneous attitude information calculation.

[0087] The instantaneous position information is calculated as follows:

[0088] In step S1, continuous functions of the three components (longitude, latitude, and altitude) of the aerial camera position were obtained through cubic spline interpolation, and are denoted as instantaneous longitude. Instantaneous latitude and instantaneous height The latitude and longitude coordinates are converted into Cartesian coordinates using the universal transverse Mercator projection, thus obtaining the instantaneous coordinates of the aerial camera center in the Cartesian coordinate system. .

[0089] The conversion formula is: ,in The region is determined by the universal transverse Mercator projection zone of the aerial survey area, derived from the design parameters of the aerial survey mission. In this embodiment... The value is 111319.99 meters. and These represent the longitude and latitude of the center of the aerial survey area corresponding to the coordinate statistics of the ground control points in step S1, and the instantaneous longitude and latitude of the camera at that moment.

[0090] Combined with airflow disturbance compensation coefficient By correcting the formula: The corrected instantaneous height was calculated. ,in , This is represented by the sampling frequency preset in step S1. This is expressed as the standard deviation of the height data after smoothing in step S1. The average height of the aerial camera during the exposure period.

[0091] The instantaneous attitude information is calculated as follows:

[0092] The continuous functions of the three components of the camera attitude (pitch angle, roll angle, and yaw angle) obtained by cubic spline interpolation in step S1 are denoted as follows: , and By correcting the formula: , , The corrected instantaneous attitude angles were calculated. , and ,in The sliding window size (dimensionless) is determined based on a preset sampling frequency. and exposure duration Using the formula: The sliding window is calculated. In this embodiment Take 1, The time interval preset in step S1 This is represented as the attitude coupling correction coefficient. This represents the sampling point number within the sliding window, used to iterate through and accumulate the attitude angle data at different times within the window.

[0093] Based on the corrected instantaneous camera position and attitude angle, combined with the aerial camera intrinsic parameters extracted in step S1, including focal length... Pixel size Based on the principal point offset, the instantaneous motion trajectory of each pixel within the image plane of each frame is constructed, and the displacement change of the pixel during the exposure period is quantified and calculated, specifically as follows: ,in For a moment Pixels The instantaneous offset within the image plane. Represented as time Pixels The instantaneous offset of the horizontal coordinate in the image plane. Represented as time Pixels The instantaneous offset of the ordinate in the image plane. , , and These are respectively represented as the exposure start time point (i.e. The horizontal and vertical coordinates of the center of the aerial camera.

[0094] All instantaneous moments within the exposure period of Integrate to obtain the pixel. Complete motion trajectory throughout the entire exposure period By traversing all pixels in the plane corresponding to each frame of the image, a pixel-level motion trajectory model of each frame of the image is constructed.

[0095] It should be noted that the value of the attitude coupling correction coefficient is determined by statistical quantification of the measured attitude data after preprocessing in step S1. First, linear correlation analysis is performed on multiple sets of original pitch and roll angle data collected by the aerial survey mission to obtain the correlation coefficient α. Then, the initial value is calculated using the formula attitude coupling correction coefficient = 0.1 × α, and it is constrained within a reasonable range of 0.01 to 0.05. In this embodiment, the correlation coefficient α between pitch and roll angles is 0.3 obtained from the statistical analysis of measured data. Therefore, the attitude coupling correction coefficient is 0.03. This value is determined by quantification based on actual aerial survey data.

[0096] In the process of constructing a pixel-level motion trajectory model, this invention performs projection transformation and height correction on the instantaneous position of the camera, performs sliding window mean and coupling correction on the attitude angle, and calculates the instantaneous pixel offset by integrating the camera intrinsic parameters. This is beneficial for accurately restoring the true motion trajectory of pixels during exposure and accurately characterizing the distribution characteristics of non-uniform blur in aerial survey images.

[0097] In a specific embodiment, the process of generating a spatial variation blur kernel for the entire aerial survey image based on pixel-level motion trajectory analysis is as follows: for any pixel in the corresponding plane of each frame image Based on the number of pixels during the exposure period The motion trajectory is analyzed, and the total displacement amplitude of the trajectory is extracted from it. , trajectory principal direction angle Trajectory Discrete Factor .

[0098] in ,in and Represented as pixels The total offset of the horizontal and vertical axes during the exposure period.

[0099] .

[0100] , This is expressed as the exposure duration. and Represented as pixels The instantaneous lateral velocity and the instantaneous longitudinal velocity, Represents pixels lateral instantaneous offset at time The derivative at that point, i.e., the pixel At any moment The small lateral displacement increment, Similarly.

[0101] Based on the above trajectory characteristics, namely the total displacement amplitude of the trajectory , trajectory principal direction angle Trajectory Discrete Factor A spatially varying blur kernel is generated for each pixel.

[0102] Where the fuzzy kernel size Fuzzy kernel weights , Indicates the current pixel In the corresponding spatially varying fuzzy kernel, the kernel coordinates The weight value at each location (dimensionless, ranging from 0 to 1, with the sum of the weights of all locations within the kernel being 1). This represents the time sampling interval of the pixel motion trajectory corresponding to the coordinates within the blur kernel. Represents pixels instantaneous moment The trajectory discrete factor.

[0103] After the kernel size and weight calculations are completed, the direction mentioned is the principal direction angle of the pixel motion trajectory. The corresponding fuzzy extension direction is consistent with the actual motion direction during the aerial survey camera exposure, which can distinguish forward motion, lateral drift, and irregular motion directions caused by airflow disturbances. Direction generation in pixels A spatial variation fuzzy kernel centered on and matching the corresponding fuzzy extension trend. This enables refined modeling of each frame of image, with one pixel and one kernel corresponding to each frame.

[0104] After constructing the blur kernel by traversing all pixels of each frame of the image, the following calculation formula is used: Calculate the global blur intensity factor corresponding to a single frame image. , and These represent the length and width of a single frame of the image, respectively.

[0105] It should be noted that, In the formula, coefficient 2 is used to symmetrically expand the pixel motion trajectory on both sides, ensuring that the blur kernel can completely cover the entire motion range of the pixel during exposure, thus achieving bilateral symmetrical modeling of the blur trajectory. Constant 1 is used to ensure that the final calculated blur kernel size is odd, so that the blur kernel has a unique center pixel, satisfying the core requirement of image convolution operation. This combination of coefficients is a fixed structural parameter determined by this scheme for the symmetrical distribution characteristics of aerial survey motion blur trajectory. It does not need to be dynamically adjusted according to the scene, but only the displacement amplitude after rounding up is used. It can adaptively determine the optimal kernel size to suit the current pixel blur level.

[0106] With the current processing pixel As the center of the fuzzy kernel, based on the size of the fuzzy kernel Define the pixel coverage area of ​​the kernel, and then use the principal direction angle of the pixel motion trajectory. As the main axis of fuzzy extension, the coordinates inside the fuzzy kernel A rotational transformation mapping is performed along the principal axis to align it with the blur direction formed by the actual camera motion, followed by weight normalization according to the trajectory density. Assign values ​​to each position within the kernel and perform overall weight normalization, ultimately generating a result with... Centered on, along Directional extension and spatial variation fuzzy kernel that conforms to the motion fuzzy characteristics of aerial surveying .

[0107] In the process of generating a spatially varying fuzzy kernel, this invention extracts trajectory parameters based on the total offset of pixel horizontal and vertical coordinates, the direction of motion velocity, and the distribution of displacement increments. Then, based on the trajectory features, it determines the size, principal axis, and weight of the fuzzy kernel. This approach helps to abandon the traditional globally unified fuzzy kernel and achieve pixel-by-pixel adaptive fuzzy modeling.

[0108] S3. Based on image overlap, candidate frames are extracted, and adjacent images that match the frame space to be processed are selected. The quality of each candidate frame is calculated based on parameters such as the fuzzy intensity factor in step S2, and threshold discrimination and optimal frame subset are determined. After removing inferior frames, an effective frame set suitable for subsequent fusion is obtained.

[0109] In a specific embodiment, the initial extraction process of candidate frames based on camera pose and ground projection is as follows: based on the aerial camera center height of the single-frame image in step S1... ,focal length Image size and pixel size Using the formula: The total ground projection area of ​​a single frame image is calculated. .

[0110] Next, project the four corner points of the plane of the frame to be processed and the corresponding images of adjacent frames onto the ground coordinate system, and solve for the intersection area of ​​the projected polygons. Using the formula: The ground projection overlap rate corresponding to a single frame image is calculated. .

[0111] Combined with the global blur intensity factor corresponding to the single frame image in step S2 Set an adaptive overlap threshold : Filter out The adjacent frames are used to form a candidate frame set.

[0112] It should be noted that, based on the camera intrinsic parameters, POS position coordinates, and attitude angle parameters obtained in step one, the four corner points of the image plane of the frame to be processed and the adjacent frames are projected point by point onto the same local ground rectangular coordinate system using the photogrammetric collinearity equation, forming two ground convex quadrilateral projection polygons. Then, the Sutherland-Hodgman polygon clipping algorithm is used to find the overlapping closed area of ​​the two convex quadrilaterals. Finally, the area of ​​the overlapping closed area is calculated using the shoelace formula, which is the ground projection intersection area of ​​the two frames. For example, after a set of adjacent aerial survey frames are projected onto the ground, each forming a 50m×40m convex quadrilateral, the intersection area is obtained by clipping and finding the regular rectangular overlapping area. Substituting into the shoelace formula, the intersection area is directly calculated to be 1200m².

[0113] The adaptive overlap threshold of 0.65 represents the threshold for aerial imagery to be blur-free (i.e., ...). The baseline overlap threshold in the =0 state is determined empirically by conventional aerial survey heading / lateral optimal overlap, ensuring the selection of candidate frames with the best spatial complementarity under clear imagery; 0.2 is the blur intensity adaptive adjustment coefficient, used to adjust the blur intensity factor. The overlap threshold is dynamically lowered. This value is determined by experimental fitting of multiple sets of aerial survey images with different degrees of blur, taking into account both the effectiveness of the screening and the rationality of the scope. In practical applications, based on the platform, flight altitude and shooting conditions of the target aerial survey area, the baseline value and adjustment coefficient are slightly calibrated and adapted through correlation analysis between the blur intensity and effective overlap of the sample images.

[0114] In the candidate frame extraction process, this invention calculates the ground projection overlap rate using camera intrinsic parameters and flight altitude, and constructs an adaptive overlap threshold by combining the fuzzy intensity factor. This helps to eliminate adjacent frames with insufficient overlap and poor matching, thus ensuring the effective fusion value of candidate frames.

[0115] In a specific embodiment, the process of multi-dimensional quality quantization scoring of candidate frames combined with fuzzy priors is as follows:

[0116] The candidate frame set selected in step S3, the global blur intensity factor corresponding to each frame image in step S2, the total displacement amplitude of the pixel-by-pixel trajectory, and the image residual preprocessed in step S1 are used as inputs.

[0117] Combined with the length corresponding to a single frame image ,Width Pixels corresponding to a single frame image Total displacement amplitude of the trajectory during the exposure period Through the formula: The average blur displacement amplitude of the candidate frames was calculated. .

[0118] Filter the maximum average blur displacement amplitude of a single frame image in the candidate frame set. Through the formula: The blur fit score corresponding to a single frame image is calculated. .

[0119] Combining spatial gradient entropy and frequency domain high-frequency energy entropy and fixed equilibrium coefficient Through the formula: The spatial-frequency joint sharpness entropy corresponding to a single frame image is calculated. .

[0120] Combined with the corresponding image average noise residual in the candidate frame set (Derived from the residual after denoising in step S1) and the corresponding maximum average noise residual in the candidate frame set. Through the formula: The noise suppression score corresponding to a single frame image is calculated. .

[0121] Combining the blur fit score corresponding to a single frame image The joint sharpness entropy of spatial and frequency domains corresponding to a single frame image Noise suppression score corresponding to a single frame image And by combining the preset weighting factors, a comprehensive quality score for the corresponding single frame image in the candidate frame set is obtained through weighted calculation.

[0122] It should be noted that the spatial gradient entropy is obtained by calculating the pixel-by-pixel gradient magnitude of a single-frame grayscale image after preprocessing in step S1 using the Sobel operator, statistically analyzing the normalized histogram probability distribution of the gradient magnitude, and then solving it according to the information entropy formula; the frequency domain high-frequency energy entropy is obtained by performing a two-dimensional Fourier transform on the image frame, separating the high-frequency components, calculating their energy distribution, and then obtaining it using the information entropy formula, which is used to characterize the richness of high-frequency details in the frequency domain.

[0123] The fixed balance coefficient was determined through a large number of experimental aerial survey images: First, an aerial survey image sample set containing different degrees of blur, texture complexity, and flight conditions was collected. The fixed balance coefficient was then used to conduct multiple comparative tests from 0 to 1 in steps of 0.1. The accuracy and robustness of the joint sharpness entropy in distinguishing between sharp and blurry states of the images were calculated under different values. When the fixed balance coefficient was 0.5, the representation weights of spatial gradient information and high-frequency energy in the frequency domain were balanced, resulting in the best discrimination effect and the strongest stability for various types of motion blur and airflow disturbance blur. Therefore, it was set as a fixed value and did not need to be dynamically adjusted according to the scene.

[0124] The pre-set weighting factors were determined through comparative experiments and fitting calibration of a large number of aerial survey image samples. This process is based on existing multi-dimensional index weight optimization methods, which will not be elaborated here.

[0125] In the multi-dimensional quality scoring process, the embodiments of the present invention use fuzzy adaptation, spatial-frequency joint sharpness entropy, and noise suppression score weighted calculation to objectively quantify the quality of a single frame image and provide a scientific basis for fusion weight allocation. In the optimal fusion frame selection process, the adaptive quality threshold and quantity constraint range are dynamically adjusted to ensure a stable number of fusion frames and optimal quality, avoiding insufficient or redundant frames.

[0126] In a specific embodiment, the process of adaptive quality threshold discrimination and optimal fusion frame subset determination is as follows: a comprehensive quality score threshold is constructed by combining the clear image baseline quality threshold, the blur adaptive adjustment coefficient, and the blur intensity factor.

[0127] Set an optimal frame quantity constraint range, and filter out candidate frames in the candidate frame set whose comprehensive quality score is greater than or equal to the comprehensive quality score threshold. If the number is insufficient, the threshold is relaxed to make up for it. If it exceeds the limit, the frame with the highest comprehensive quality score is selected (the upper limit of the constraint range).

[0128] It should be noted that the blur intensity factor of the current frame to be processed, calculated in step S2, is used as the basis for dynamic adjustment. A baseline quality threshold of 0.7 under clear image conditions is used as the base value. The adaptive threshold reduction is obtained by multiplying the blur intensity factor by the blur adaptive adjustment coefficient of 0.3. Thus, the adaptive comprehensive quality score threshold is constructed to be equal to 0.7 − 0.3 × The threshold will adaptively decrease as the image blur increases. Then, the candidate frames are sorted in descending order of the comprehensive quality score Q, and frames with a score greater than or equal to the comprehensive quality score threshold are selected. If the number of frames that meet the condition is less than 3, the threshold is slightly lowered by a fixed step size to make up the minimum number of frames. If the number of frames that meet the condition is more than 5, the top 5 frames with the highest scores are directly selected, and finally a fusion frame subset with stable quantity and optimal quality is obtained.

[0129] The optimal frame count constraint range is determined based on engineering experience and extensive experimental calibration of multi-frame fusion deblurring of aerial survey images. For example, 3 frames is the minimum effective number of frames to ensure the fusion deblurring effect, and 5 frames is the optimal upper limit of frames to balance computational efficiency and fusion accuracy. This value is based on the existing multi-frame fusion optimization method and will not be elaborated here.

[0130] The "relaxed threshold supplement" refers to reducing the quality threshold by a fixed small step size when the number of selected frames that meet the threshold is less than the minimum number of valid frames (3 frames). Candidate frames are then re-selected until the minimum number of valid frames is met, ensuring that there is sufficient data to support subsequent fusion and deblurring.

[0131] S4. The optimal fusion frame subset selected is partitioned and aligned for fusion based on pixel-by-pixel blur prior, and the initial fusion image is optimized for adaptive noise reduction and detail enhancement to output the image after sharpening.

[0132] In a specific embodiment, the process of partitioning and aligning the selected optimal fusion frame subset based on pixel-by-pixel blur prior is as follows: based on the average blur displacement amplitude of the candidate frames... The calculation formula is used to calculate the average blur displacement amplitude corresponding to the optimal fused frame subset, which is then used as the global threshold for the current frame. .

[0133] Combination Pixels corresponding to a single frame image Total displacement amplitude of the trajectory during the exposure period ,when > When the blurriness is high, it is determined to be a heavily blurred region; otherwise, it is determined to be a lightly blurred region.

[0134] in A global scalar threshold calculated for the entire current frame to be processed. The point-by-point values ​​corresponding to each pixel within the frame are both derived from the same frame image and have completely consistent dimensions. The values ​​are compared with the global unified threshold within the same frame to determine whether the pixel belongs to a lightly blurred area or a heavily blurred area.

[0135] Lightly blurred regions are fused using quality-weighted fusion, specifically:

[0136] By combining the comprehensive quality scores of the corresponding single-frame images in the candidate frame set, the comprehensive quality score of each fused frame is obtained. The comprehensive quality scores of each fused frame are summed to obtain the total comprehensive quality score of the fused frame. The comprehensive quality score of each fused frame is divided by the total comprehensive quality score of the fused frame to obtain the fusion weight of each fused frame. Combining the gray values ​​of each pixel corresponding to each fused image, the fusion weight of each fused frame and the gray values ​​of each pixel corresponding to each fused image are weighted and summed to obtain the pixel gray value corresponding to each pixel after fusion.

[0137] The heavily blurred areas are blended by pixel alignment, specifically:

[0138] Total displacement amplitude of the pixel-by-pixel trajectory The corresponding pixel displacement is calculated based on the offset. Then, the corresponding pixels in each candidate frame are offset and aligned at the pixel level according to the pixel displacement. Finally, the arithmetic mean of the aligned pixel values ​​of multiple frames is calculated to obtain the fused pixel value of the heavily blurred area.

[0139] Finally, the fusion results of the lightly blurred area and the heavily blurred area are stitched together according to pixel position to output the initial fused and deblurred image.

[0140] It should be noted that the candidate frame set is all the image frames to be optimized obtained from the initial screening. Each fusion frame is the optimal fusion frame subset selected from the candidate frame set based on the comprehensive quality score and through an adaptive quality threshold. Only this subset participates in the subsequent multi-frame alignment fusion deblurring process.

[0141] For pixel-level fusion in heavily blurred regions, the offset is based on the total displacement amplitude of the pixel trajectory. First, the pixel displacement amount and displacement direction corresponding to each pixel in the image are solved. For example, if the total displacement amplitude of a certain pixel trajectory is 6 pixels and the displacement direction is along the corresponding main direction angle of the trajectory, then the corresponding position pixels of each candidate frame are offset and aligned by 6 pixels along the displacement direction according to the displacement parameter. The corresponding pixel grayscale values ​​after alignment of 3 frames in the optimal fusion frame (such as 142, 146 and 144) are selected, and the arithmetic mean of the pixel values ​​of multiple frames is calculated, that is, (142+146+144) / 3=144, so as to obtain the fused pixel value of the pixel in the heavily blurred region.

[0142] In a specific embodiment, the process of outputting the sharpened image is as follows: combining the basic noise reduction coefficient calibrated through aerial survey image experiments. and noise reduction adjustment coefficient and combined with fuzzy intensity factor Through the formula: The noise reduction coefficient was calculated. The greater the blur intensity, the lower the noise reduction intensity adaptively, avoiding excessive smoothing that could lead to the loss of blurry areas and the restoration of details. After noise reduction is completed, the noise-reduced image is output.

[0143] Combined with basic enhancement coefficient and fuzzy intensity factor Through the formula: The detail enhancement coefficient was calculated. The final output is a clear, optimized aerial survey image with no blur, low noise, and high detail.

[0144] It should be noted that both the base noise reduction coefficient and the noise reduction adjustment coefficient were obtained through comparative experiments and calibration using multiple sets of measured aerial survey image samples covering different blur intensities, noise levels, flight conditions, and ground scenes. In the experiments, the core optimization objectives were the amount of noise residue after image denoising, the integrity of details preserved in heavily blurred areas, and the absence of artifacts. The two coefficients were tested by traversing the range of reasonable values, and the optimal fixed value was selected as the final value, which enables the adaptive noise reduction module to fully suppress noise in lightly blurred areas and avoid excessive smoothing and loss of repaired details in heavily blurred areas. This value is suitable for the entire process of aerial survey image deblurring in this solution.

[0145] The basic enhancement coefficient was obtained through comparative experiments using multiple sets of measured aerial survey image samples covering different blur intensities, surface texture complexity, flight attitude disturbances, and lighting conditions. The experiments focused on edge sharpness, texture fidelity, and the absence of artifacts and over-enhancement after image detail enhancement as core optimization indicators. The basic enhancement coefficient was tested by traversing a reasonable range of preset values. The optimal fixed value was selected as the final value, which enables the dynamic detail enhancement module to effectively enhance details in lightly blurred areas, moderately enhance details in heavily blurred areas, and not produce artifacts. This value is suitable for the entire process of aerial survey image deblurring in this solution.

[0146] When optimizing the initial fused and deblurred image, the global blur intensity factor of the image and the frame to be processed, the preprocessed image noise residual, and the preset basic enhancement coefficient are used as inputs. First, an adaptive noise reduction intensity coefficient is calculated based on the image noise residual and the global blur intensity factor. For example, when the global blur intensity factor is close to 0 in a lightly blurred area, a smaller value of the noise reduction coefficient is used to perform weak noise reduction processing to preserve the original details of the image to the greatest extent. When the factor is close to 1 in a heavily blurred area, a larger value of the noise reduction coefficient is used to perform strong noise reduction processing to effectively suppress the noise accompanying the blurred area. Then, a dynamic detail enhancement coefficient is calculated based on the global blur intensity factor and the preset basic enhancement coefficient. A larger enhancement coefficient is used in lightly blurred areas to enhance edge and texture details, while a smaller enhancement coefficient is used in heavily blurred areas to gently enhance details, avoiding artifacts and noise amplification caused by excessive enhancement. Through regional adaptive adjustment of the noise reduction coefficient and the detail enhancement coefficient, an optimized aerial survey deblurred image with low noise, clear details, and no artifacts is finally obtained.

[0147] In the partition alignment fusion process of this invention, light and heavy blur regions are divided by a blur displacement threshold, and rigid alignment weighted fusion and pixel-level offset alignment mean fusion are adopted respectively. This is beneficial to both preserve details in light blur regions and eliminate ghosting in heavy blur regions, thereby improving fusion accuracy. In the post-image optimization process, noise reduction and enhancement coefficients are dynamically calculated by blur intensity factor, which is beneficial to achieve regional adaptive processing, avoid loss of details, artifacts and noise amplification, and significantly improve the clarity and fidelity of aerial survey images.

[0148] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement any one of the steps of an aerial survey image deblurring method using a multi-frame fusion algorithm.

[0149] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for deblurring aerial survey images based on a multi-frame fusion algorithm, characterized in that, include: The location and attitude information of the aerial survey platform and the aerial survey image sequence are acquired. Data preprocessing and spatiotemporal alignment are performed to obtain the continuous camera motion trajectory function, the preprocessed image sequence and camera intra-camera parameters. The location information includes longitude, latitude, and altitude, and the attitude information includes pitch angle, roll angle, and heading angle; Extract position and attitude information within the exposure time interval of a single frame of aerial survey imagery, construct a pixel-level motion trajectory model on the image plane, analyze and generate the spatial variation blur kernel of the entire aerial survey imagery, and calculate the global blur intensity factor. Candidate frame set is extracted based on the overlap of image ground projection. The candidate frames are then subjected to multi-dimensional quality quantification scoring. The optimal fused frame subset is selected by adaptive quality threshold discrimination and optimal frame quantity constraint. The global average blur displacement threshold is calculated based on the total displacement amplitude of the pixel-by-pixel trajectory of the current frame to be processed. The optimal fusion frame subset is then fused and aligned according to the light and heavy blur regions to obtain the initial fused deblurred image. Based on image noise characteristics and global blur intensity factor, the initial fused image is optimized for noise reduction and detail enhancement, and finally the cleared aerial survey image is output.

2. The aerial survey image deblurring method based on a multi-frame fusion algorithm according to claim 1, characterized in that, Through data preprocessing and spatiotemporal alignment, we obtain the continuous camera motion trajectory function, the preprocessed image sequence, and the camera's intra-camera parameters, including: Outlier removal and Kalman filtering smoothing are performed on the position and attitude data collected by the aerial survey platform's positioning and attitude determination system. The continuous camera motion trajectory function is then obtained by cubic spline interpolation fitting. Radiometric correction and nonlocal mean denoising were performed sequentially on the original aerial survey images to extract the camera parameters and exposure time intervals of each frame. Ground control points within the aerial survey area are selected, and the time offset is optimized with the goal of minimizing the sum of squared residuals between the projected coordinates of the control points and the actual observed coordinates. The timestamps of the motion data are unified to the shutter time reference of the aerial camera to complete the spatiotemporal alignment.

3. The aerial survey image deblurring method based on a multi-frame fusion algorithm according to claim 2, characterized in that, Constructing a pixel-level motion trajectory model within the image plane specifically includes: The exposure time range for each frame of the image is defined based on the start and end times of the exposure. By calculating instantaneous position and attitude information, the instantaneous position and attitude information of the aerial camera at each moment within the exposure time interval are obtained. Based on the instantaneous longitude, instantaneous latitude, and instantaneous altitude of the aerial camera's position, the latitude and longitude coordinates are converted into Cartesian coordinates to obtain the instantaneous coordinates of the aerial camera's center in the Cartesian coordinate system. By combining the airflow disturbance compensation coefficient and the instantaneous height, the corrected instantaneous height, i.e., the instantaneous position information, is calculated; Instantaneous pitch, roll, and yaw angles of the camera attitude obtained through cubic spline interpolation; By combining the preset sampling frequency and exposure time interval, a sliding window is calculated. Combined with the attitude coupling correction coefficient and the preset time interval, the corrected instantaneous attitude information is calculated. Based on the corrected instantaneous position and attitude angle of the camera, combined with the aerial camera's intrinsic parameters, including focal length, pixel size and principal point offset, the instantaneous motion trajectory of each pixel in the image plane of each frame is constructed. Integrating over all instants during the exposure period yields the complete motion trajectory of a single pixel throughout the entire exposure. By traversing all pixels in the corresponding plane of each frame, a pixel-level motion trajectory model for each frame is obtained.

4. The aerial survey image deblurring method based on a multi-frame fusion algorithm according to claim 3, characterized in that, The process includes generating a spatially varying fuzzy kernel and calculating the global fuzzy intensity factor, including: For each pixel in a single frame image, the total displacement amplitude of the trajectory is extracted by synthesizing and calculating the total offset of the horizontal and vertical coordinates of that pixel during the exposure period. Based on the directional relationship between the horizontal and vertical instantaneous motion velocities of the pixel during the exposure period, the main trajectory direction angle is extracted. The trajectory discrete factor is extracted based on the statistical calculation of the displacement increment distribution within the trajectory time sampling interval during the exposure period. The size of the fuzzy kernel is determined by symmetric expansion and rounding operations based on the total displacement amplitude of the trajectory. The main direction angle of the trajectory is used as the main axis of fuzzy extension and the coordinates in the kernel are rotated and mapped. The weight distribution in the kernel is calculated by combining the trajectory discrete factor and the weight normalization is completed. A spatial variation fuzzy kernel that fits the motion fuzzy characteristics of aerial surveying is generated with the current pixel as the center. The total displacement amplitude of the trajectory of all pixels in the whole frame is statistically analyzed and normalized to obtain the global blur intensity factor of the single frame image.

5. The aerial survey image deblurring method based on a multi-frame fusion algorithm according to claim 4, characterized in that, Candidate frame sets are extracted based on image overlap, including: Obtain the center height of the aerial camera, the focal length, image size and pixel size in the camera's intrinsic parameters corresponding to a single frame image, and calculate the total ground projection area corresponding to each frame image. Project the four corner points of the image of the frame to be processed and the adjacent frame onto the same ground rectangular coordinate system to form a ground projection polygon and calculate the intersection area of ​​the two. Combine the total ground projection area and the intersection area to calculate the ground projection overlap rate of the two images. An adaptive overlap threshold is constructed by combining the global blur intensity factor of a single frame image, and adjacent frames that meet the overlap threshold requirements are selected to form an initial candidate frame set.

6. The aerial survey image deblurring method based on a multi-frame fusion algorithm according to claim 5, characterized in that, Multi-dimensional quality quantization scoring of candidate frames is performed by combining fuzzy priors, including: The inputs are the candidate frame set, the global blur intensity factor of each frame image, the total displacement amplitude of the pixel-by-pixel trajectory, and the preprocessed image noise residual. The average blur displacement amplitude of the candidate frame is calculated by combining the image size and the total displacement amplitude of the pixel-by-pixel trajectory. Based on this average blur displacement amplitude, the blur fit score of a single frame image is calculated. By combining the spatial gradient entropy, the frequency high-frequency energy entropy, and the fixed balance coefficient, the joint spatial-frequency sharpness entropy of a single frame image is obtained. The noise suppression score of a single frame image is obtained by combining the average noise residual of the candidate frame set and the maximum average noise residual. The fuzziness fit score, spatial-frequency joint sharpness entropy, and noise suppression score are weighted and summed using preset weighting factors to obtain the comprehensive quality score of a single frame image in the candidate frame set.

7. The aerial survey image deblurring method based on a multi-frame fusion algorithm according to claim 6, characterized in that, Quality threshold discrimination and optimal fusion frame subset determination include: By combining the clear image baseline quality threshold, the blur adaptive adjustment coefficient, and the global blur intensity factor of the frame to be processed, a comprehensive quality score threshold that adjusts with the degree of blur is constructed. Set a constraint range for the optimal number of fused frames, sort the candidate frames according to the comprehensive quality score, and select candidate frames with a comprehensive quality score ≥ the comprehensive quality score threshold. If the number of qualified candidate frames is lower than the lower limit of the constraint interval, the overall quality score is reduced by a fixed step size and the number of frames is re-selected to make up the difference. If the number of candidate frames that meet the conditions is higher than the upper limit of the constraint interval, then the upper limit of the constraint interval is used as the selection criterion, and the frame image with the highest comprehensive quality score is selected to obtain the optimal fused frame subset.

8. The aerial survey image deblurring method based on a multi-frame fusion algorithm according to claim 7, characterized in that, Calculate the global average blur displacement threshold and perform partition alignment fusion on the optimal fused frame subset, including: The global average blur displacement threshold of the current frame to be processed is obtained by performing a global statistical average calculation based on the total displacement amplitude of the trajectory of all pixels in the frame to be processed. Using the global average blur displacement threshold as the region discrimination criterion, the entire image domain of the frame to be processed is divided into lightly blurred regions and heavily blurred regions. For lightly blurred areas, the fusion weight of each fused frame is calculated by combining the comprehensive quality score of the corresponding single frame image in the candidate frame set, and the pixel gray value of each pixel after fusion is calculated by combining the gray value of each pixel corresponding to each frame fused image. For heavily blurred regions, the corresponding pixel displacement is calculated based on the total displacement amplitude of the pixel trajectory. Then, the corresponding pixels in each candidate frame are aligned with pixel-level offsets based on the pixel displacement. Finally, the arithmetic mean of the aligned pixel values ​​of multiple frames is calculated to obtain the fused pixel value of the heavily blurred region. For the optimal fused frame subset after partition alignment, corresponding weighted fusion processing is performed on the lightly blurred and heavily blurred regions respectively, and the initial fused deblurred image is finally obtained.

9. The aerial survey image deblurring method based on a multi-frame fusion algorithm according to claim 8, characterized in that, Adaptive noise reduction and dynamic detail enhancement are performed on the initial fused image, including: The initial fused deblurred image, the global blur intensity factor of the frame to be processed, the preprocessed image noise residual, and the preset basic enhancement coefficient are used as inputs to perform adaptive noise reduction processing first. By combining the noise residual of the initial fused image with the global blur intensity factor of the frame to be processed, an adaptive noise reduction intensity coefficient that dynamically adjusts with the degree of blur is calculated, and the initial fused image is subjected to regional noise reduction processing based on this coefficient. Then, dynamic detail enhancement processing is performed. The dynamic enhancement coefficient is calculated by combining the global blur intensity factor of the frame to be processed with the preset basic enhancement coefficient. Based on this coefficient, the details of the denoised fused image are enhanced. After noise reduction and detail enhancement optimization, the aerial survey images are output with improved clarity.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it performs the steps of any one of claims 1 to 9.

Images