An unmanned aerial vehicle aerial photograph image optimization method and system based on artificial intelligence

Abstract

The present application relates to the technical field of image optimization, in particular to a UAV aerial image optimization method and system based on artificial intelligence, comprising the following steps: calling flight state information to analyze attitude deviation, calculating the sun irradiation angle to adjust exposure, evaluating noise indicators to adjust filtering, identifying scene labels to match processing parameters, generating gradient heat maps to optimize texture, and obtaining aerial image optimization records. In the present application, by calling real-time UAV flight attitude data and light change trend, the attitude and light state deviation in the shooting process are finely evaluated, the image geometry and light parameters are precisely adjusted, the filtering strength is dynamically optimized by using the regional noise deviation amplitude, the scene features are constructed according to the local semantic density, the texture reconstruction strength is dynamically classified and optimized according to the local gradient change of the image, the problems of uneven illumination and regional detail loss of aerial images are effectively reduced, and the overall visual clarity and content recognition accuracy of the image are improved.

Description

Technical Field

[0001] This invention relates to the field of image optimization technology, and in particular to an artificial intelligence-based method and system for optimizing drone aerial images. Background Technology

[0002] Image optimization technology encompasses methods such as digital image processing, pattern recognition, and image analysis and enhancement to improve and adjust image content, enhancing visual quality such as clarity, contrast, and color reproduction, or adapting to the needs of subsequent recognition and application processing. It includes multiple technical directions such as image preprocessing, noise suppression, image enhancement, edge detection, target extraction, and image reconstruction, covering rule-based image transformation operations, including image semantic understanding and content adaptive adjustment achieved through neural networks and deep learning. It is applied in scenarios such as remote sensing monitoring, security identification, medical imaging, and industrial inspection, and has become a key link in UAV aerial photography, video analysis, and intelligent recognition. Among these, AI-based UAV aerial image optimization methods refer to methods for optimizing aerial images acquired by UAVs. To address issues such as image blurring, uneven lighting, and missing details in target areas, convolutional neural networks are used for image denoising, enhancement, and feature reconstruction. Combined with an image super-resolution reconstruction model, low-quality images are upgraded in resolution. Furthermore, an attention-based mechanism for judging the importance of image regions is employed to automatically identify and adjust the clarity and contrast of key regions in the image, optimizing the image content structure at both the pixel and semantic levels. The method introduces a collaborative correction mechanism between UAV flight parameters and image metadata. By combining information such as GPS track, heading angle, and shooting time, an image quality prediction model is constructed, and appropriate optimization parameter configuration paths are selected. This enables targeted adjustment of image region distribution and automatic scheduling of processing instructions within the image optimization process.

[0003] Traditional UAV aerial image optimization technology focuses on a unified and standardized processing flow in practical applications. The image optimization parameters are relatively fixed and lack a real-time dynamic adjustment mechanism. It fails to fully and accurately perceive and adapt to the real-time changes in lighting conditions, flight attitude, and local details of aerial images. This leads to problems such as loss of local details and uneven lighting when the flight environment is complex or the local area is abnormal, which affects the actual effect of aerial images in detail presentation and subsequent intelligent recognition applications. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an artificial intelligence-based method and system for optimizing drone aerial images.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an artificial intelligence-based method for optimizing drone aerial images, comprising the following steps:

[0006] S1: Call the navigation status information, analyze the continuous change trajectory of the UAV's pitch angle, roll angle and heading angle in the time series, assess the degree of attitude deviation of the UAV, and obtain the set of geometric adjustment parameters based on the angle change direction and speed information;

[0007] S2: Call the geometric adjustment parameter set, call the shooting timestamp, calculate the angle range between the sun's illumination direction and the shooting field of view during image acquisition, evaluate the balance of exposure conditions, adjust the image preprocessing configuration, and obtain the illumination compensation parameter set;

[0008] S3: Call the illumination compensation parameter set, calculate the brightness change frequency, edge texture jitter and gray scale distribution fluctuation rate in the image block area, construct a noise index set, average the index of each region to form a full-image reference index set and compare it with each region, evaluate the noise offset amplitude of each region and adjust the noise filtering intensity to obtain the enhancement adjustment configuration;

[0009] S4: Using the enhanced adjustment configuration, analyze the color distribution discreteness, boundary direction aggregation and texture concentration in the image block region, construct a semantic density vector, identify scene labels and configure image processing parameters, and obtain scene configuration matching results.

[0010] As a further embodiment of the present invention, the geometric adjustment parameter set specifically includes rotation angle configuration data, cropping boundary coordinate information, and repositioning pixel center data; the illumination compensation parameter set includes exposure control level, photosensitivity response coefficient, brightness adjustment parameter, and contrast enhancement factor; the enhancement adjustment configuration specifically includes filter response adjustment level, regional noise suppression ratio, and image enhancement control factor; and the scene configuration matching result includes image spatial label set, regional semantic recognition number, and processing template selection mark.

[0011] As a further aspect of the present invention, the step of obtaining the geometric adjustment parameter set specifically includes:

[0012] S111: Call the navigation status information to obtain continuous time series data of the UAV's pitch angle, roll angle and heading angle, analyze the angle change direction and velocity change status of each sampling time period, and establish an angle change dataset;

[0013] S112: Based on the angle change dataset, calculate the range of change of each angle in a continuous time period, calculate the attitude stability coupling coefficient, and output the attitude offset dynamic index.

[0014] S113: Based on the attitude offset dynamic index, combined with the angle change direction and flight speed trend, configure the adjustment parameters of image rotation angle, boundary cropping range and pixel centroid coordinates to generate a geometric adjustment parameter set.

[0015] As a further aspect of the present invention, the step of obtaining the illumination compensation parameter set specifically includes:

[0016] S211: Call the geometric adjustment parameter set, calculate the solar illumination direction vector and the shooting field of view direction vector based on the shooting timestamp and image acquisition location data, obtain the angle change trend of each time period, and generate the illumination angle change sequence.

[0017] S212: Based on the light angle change sequence, analyze the offset trend of the sun's illumination direction and the shooting field of view in a continuous time series, collect the corresponding lighting stability sequence and light change offset sequence, calculate the exposure balance measurement coefficient, evaluate the balance state of the exposure conditions, and establish an exposure state label.

[0018] S213: Based on the exposure status label, adjust the image preprocessing configuration according to the current scene's exposure status, including exposure time adjustment level, sensitivity control coefficient, image brightness correction intensity, and contrast enhancement ratio parameters, and generate a set of illumination compensation parameters.

[0019] As a further aspect of the present invention, the step of obtaining the enhanced adjustment configuration specifically includes:

[0020] S311: Call the illumination compensation parameter set to calculate the brightness change frequency, edge texture jitter and grayscale distribution fluctuation rate of the image block region, form the noise intensity index of each partition, and construct a noise index set;

[0021] S312: Based on the noise index set, the indices of each region of the image are averaged to form a full-image reference index set. By comparing the indices of each region with the reference indices, the comprehensive noise offset amplitude is calculated, the noise offset amplitude of each region is evaluated, and a region noise ratio matrix is ​​constructed.

[0022] S313: Based on the aforementioned regional noise ratio matrix, adjust the noise filtering intensity of each region according to the noise offset amplitude of each region to establish an enhancement adjustment configuration.

[0023] As a further aspect of the present invention, the step of obtaining the scene configuration matching result specifically includes:

[0024] S411: Using the enhanced adjustment configuration, analyze the color distribution discreteness, boundary direction aggregation and texture concentration of the image block region, construct the semantic density vector of each partition, and generate a partition semantic density set;

[0025] S412: Based on the partition semantic density set, compare the coverage and proportion trend of each semantic clustering category in the image, determine the category dominance and identify the scene label of each frame image, and obtain the scene label discrimination result;

[0026] S413: Based on the scene label discrimination result, match the corresponding image processing parameters according to the scene label of the image to obtain the scene configuration matching result.

[0027] As a further aspect of the present invention, the method further includes:

[0028] S5: Using the scene configuration matching results and flight altitude information, calculate the spatial gradient change formed by pixel projection line of sight, construct a gradient change heat map in two-dimensional space, divide the pixel region into multiple levels according to the degree of continuous span in the gradient map, adjust the texture reconstruction intensity parameters, and obtain the aerial image optimization record.

[0029] The aerial image optimization record includes depth partition mapping identifiers, region reconstruction parameter levels, and texture processing control groups.

[0030] As a further aspect of the present invention, the step of acquiring the optimized aerial image record specifically includes:

[0031] S511: Using the scene configuration matching results and the flight altitude corresponding to each image frame, calculate the ground projection line-of-sight trajectory of each pixel in the image under the current viewpoint, and obtain the pixel projection line-of-sight set.

[0032] S512: Based on the pixel projection viewing distance set, analyze the projection change span between pixels, form a spatial gradient distribution, construct a two-dimensional spatial gradient change heat map, and obtain spatial gradient heat distribution information.

[0033] S513: Based on the spatial gradient thermal distribution information, the pixel region is divided into multiple levels according to the degree of continuous span, and the texture reconstruction intensity parameters of each level are adjusted to obtain the aerial image optimization record.

[0034] An AI-based drone aerial image optimization system, wherein the AI-based drone aerial image optimization system is used to execute the aforementioned AI-based drone aerial image optimization method, the system comprising:

[0035] The attitude trajectory analysis module calls upon flight status information to analyze the continuous change trajectory of the UAV's pitch angle, roll angle, and heading angle in the time series, assesses the degree of attitude deviation of the UAV, and obtains a set of geometric adjustment parameters based on the direction and speed information of angle change.

[0036] The exposure equalization evaluation module calls the geometric adjustment parameter set, calls the shooting timestamp, calculates the angle range between the sun's illumination direction and the shooting field of view direction when the image is acquired, evaluates the equalization state of the exposure conditions, adjusts the image preprocessing configuration, and obtains the illumination compensation parameter set.

[0037] The regional noise quantization module calls the illumination compensation parameter set to calculate the brightness change frequency, edge texture jitter and grayscale distribution fluctuation rate in the image block region, constructs a noise index set, averages the index of each region to form a full-image reference index set and compares it with each region, evaluates the noise offset amplitude of each region and adjusts the noise filtering intensity to obtain the enhancement adjustment configuration.

[0038] The image semantic clustering module uses the enhanced adjustment configuration to analyze the color distribution discreteness, boundary direction aggregation and texture concentration in the image block region, constructs a semantic density vector, identifies scene labels and configures image processing parameters, and obtains scene configuration matching results.

[0039] The spatial gradient reconstruction module uses the scene configuration matching results and flight altitude information to calculate the spatial gradient change formed by the pixel projection line of sight, constructs a gradient change heatmap in two-dimensional space, divides the pixel region into multiple levels according to the degree of continuous span in the gradient map, adjusts the texture reconstruction intensity parameters, and obtains an optimized record of aerial images.

[0040] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0041] In this invention, by calling up UAV flight attitude data and illumination change trends in real time, the deviation of attitude and illumination state during the shooting process is precisely evaluated, and the image geometry and illumination parameters are accurately adjusted. The filtering intensity is dynamically optimized by utilizing the regional noise offset amplitude, scene features are constructed based on local semantic density, and the texture reconstruction intensity is dynamically and hierarchically optimized for local gradient changes in the image. This effectively reduces the problems of uneven illumination and missing regional details in aerial images, and improves the overall visual clarity and content recognition accuracy of the image. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the workflow of the present invention;

[0043] Figure 2 This is a flowchart of the process for obtaining the geometric adjustment parameter set of the present invention;

[0044] Figure 3 This is a flowchart of the illumination compensation parameter set acquisition process of the present invention;

[0045] Figure 4 This is a flowchart illustrating the enhanced adjustment configuration acquisition process of the present invention.

[0046] Figure 5 Flowchart for obtaining matching results for scenario configuration in this invention;

[0047] Figure 6 This is a flowchart of the aerial image optimization recording acquisition process of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0049] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0050] Please see Figure 1 This invention provides a technical solution: an artificial intelligence-based method for optimizing drone aerial images, comprising the following steps:

[0051] S1: Call the navigation status information, analyze the continuous change trajectory of the UAV's pitch angle, roll angle and heading angle in the time series, assess the degree of attitude deviation of the UAV, and obtain the set of geometric adjustment parameters based on the angle change direction and speed information;

[0052] S2: Call the geometric adjustment parameter set, call the shooting timestamp, calculate the angle range between the sun's illumination direction and the shooting field of view during image acquisition, evaluate the balance of exposure conditions, adjust the image preprocessing configuration, and obtain the illumination compensation parameter set.

[0053] S3: Call the illumination compensation parameter set, calculate the brightness change frequency, edge texture jitter and grayscale distribution fluctuation rate in the image block region, construct a noise index set, average the index of each region to form a full-image reference index set and compare it with each region, evaluate the noise offset amplitude of each region and adjust the noise filtering intensity to obtain the enhancement adjustment configuration;

[0054] S4: Utilize enhanced adjustment configuration to analyze the color distribution dispersion, boundary direction aggregation, and texture concentration in image block regions, construct semantic density vectors, identify scene labels, configure image processing parameters, and obtain scene configuration matching results;

[0055] S5: Utilize scene configuration matching results and flight altitude information to calculate the spatial gradient change formed by pixel projection line of sight, construct a gradient change heatmap in two-dimensional space, divide the pixel region into multiple levels according to the degree of continuous span in the gradient map, adjust the texture reconstruction intensity parameters, and obtain an optimized record of aerial images.

[0056] The geometric adjustment parameter set specifically includes rotation angle configuration data, cropping boundary coordinate information, and repositioning pixel center data. The illumination compensation parameter set includes exposure control level, photosensitive response coefficient, brightness adjustment parameters, and contrast enhancement factor. The enhancement adjustment configuration specifically includes filter response adjustment level, regional noise suppression ratio, and image enhancement control factor. The scene configuration matching results include image spatial label set, regional semantic recognition number, and processing template selection mark. The aerial image optimization record includes depth partition mapping identifier, regional reconstruction parameter level, and texture processing control group.

[0057] Please see Figure 2 The specific steps for obtaining the geometry adjustment parameter set are as follows:

[0058] S111: Call the navigation status information to obtain continuous time series data of the UAV's pitch angle, roll angle and heading angle, analyze the angle change direction and velocity change status of each sampling time period, and establish an angle change dataset;

[0059] The navigation status information is accessed by acquiring real-time sensor data recorded by the UAV flight control system. Taking a real UAV as an example, the UAV samples flight status data 10 times per second, generating continuous time-series data, including pitch angle, roll angle, and yaw angle. Assume the flight attitude data recorded by the UAV at continuous sampling points (e.g., once every 0.1 seconds) is as shown below, as shown in Table 1:

[0060] Table 1. Flight Attitude Data of Unmanned Aerial Vehicles:

[0061]

[0062] Taking pitch angle as an example, the direction of angle change at each adjacent sampling point is determined by the difference between the current angle value and the previous angle value. A positive difference indicates an increase in angle, while a negative difference indicates a decrease in angle. For example, if the pitch angle increases by 0.3° at the second sampling point compared to the first sampling point, it indicates an increase in the direction of angle change. The change value is then divided by the interval of 0.1 seconds between adjacent sampling points to obtain the rate of change within that sampling segment, which is 3° / s. This method is used to determine the direction and rate of change for each segment of angle change. The same method is applied to the analysis of roll and heading angle changes. Finally, the direction and rate of change for each segment of angle change are calculated to form a complete dataset of angle changes.

[0063] S112: Based on the angle change dataset, calculate the range of variation for each angle over a continuous time period using the formula:

[0064] ;

[0065] Calculate the attitude stability coupling coefficient and output the dynamic index of attitude deviation;

[0066] in, For the first The normalized value of the roll angle change over a time period is obtained by dividing the roll angle change for each segment by the maximum roll angle change. For the first The normalized value of pitch angle change over a time period is obtained by dividing the pitch angle change for each period by the maximum pitch angle change. For the first The normalized value of the heading angle change over a time period is obtained by dividing the heading angle change for each segment by the maximum heading angle change. For the first The normalized value of velocity change over a time period is obtained by dividing the velocity change in each segment by the maximum velocity change. For the first The normalized value of acceleration change over a time period is obtained by dividing the acceleration change in each segment by the maximum acceleration change. For the sampling point index of a continuous time series, This represents the total number of sampling points within the time series. The attitude stability coupling coefficient;

[0067] The specific process for calculating the variation range of each angle within a continuous time period using the angle change dataset is as follows: First, normalize the changes in each angle. Taking the roll angle as an example, divide each segment of change by the maximum roll angle change within the current time period to obtain the normalized value. For example, if the roll angle changes from the second sampling point to -0.2°, the third sampling point to -0.3°, and the fourth sampling point to 0.2° in Table 1, with a maximum absolute value of 0.3°, then the normalized values ​​are -0.67, -1.0, and 0.67, respectively. Similarly, the normalization of pitch and yaw angle changes is completed through similar steps to obtain the normalized value for each segment. , , Subsequently, the speed and acceleration data recorded by the drone were normalized. The speed changes of the drone at four consecutive sampling times were set to 0.2 m / s, 0.3 m / s, and 0.1 m / s, with a maximum change of 0.3 m / s. Therefore, the normalized speed change values ​​were 0.67, 1.0, and 0.33, respectively. If the acceleration changes were 0.5 m / s², 0.4 m / s², and 0.3 m / s², the maximum value was 0.5 m / s², and the corresponding normalized values ​​were 1.0, 0.8, and 0.6.

[0068] Finally, the normalized value is substituted into the formula for attitude stability coupling coefficient for calculation:

[0069] ;

[0070] Input data:

[0071] ;

[0072] ;

[0073] ;

[0074] ;

[0075] The attitude stability coupling coefficient is a dimensionless value used to quantitatively reflect the overall change level of attitude angles (roll, pitch, and yaw) under unit speed and acceleration disturbances during UAV aerial photography. A larger value indicates a greater amplitude of attitude angle change under the same flight dynamic disturbances, more significant instability at the time of acquisition, and greater susceptibility to motion blur, tilt, and distortion in the aerial images. A smaller value indicates a more stable attitude angle under flight dynamic disturbances, with lower risk to image acquisition quality. This coefficient serves as a dynamic criterion for image acquisition stability screening and dynamic geometric correction parameter configuration, and is a core reference quantity for end-to-end dynamic quality control and subsequent image optimization processes. The results show that the attitude stability coupling coefficient is 0.533, falling within the range of 0 to 1. A value closer to 0 indicates greater attitude stability, while a value closer to 1 indicates more severe attitude fluctuations.

[0076] S113: Based on the attitude offset dynamic index, combined with the angle change direction and flight speed trend, configure the adjustment parameters of image rotation angle, boundary cropping range and pixel centroid coordinates to generate a geometric adjustment parameter set;

[0077] The obtained attitude stability coupling coefficient result is 0.533. Combining the angle change direction and flight speed trend, the process of adjusting the parameters of image rotation angle, boundary cropping range and pixel centroid coordinates is as follows: The coefficient result is compared with the preset attitude stability threshold (taken as 0.6). The current calculation result is less than the threshold, indicating that the attitude is relatively stable, but there is some fluctuation. Therefore, the image rotation angle is set to a small range for correction. For example, the ratio of the attitude stability coefficient to the threshold difference (i.e., 0.6-0.533=0.067) corresponds to a scaling factor of 0.067×5°, so the image rotation angle is adjusted to 0.335°. For the image boundary cropping range, the cropping range is set according to the correlation between flight speed (e.g., 5m / s) and attitude stability. A wider boundary (0.5%~1.5% of image width) is used at higher speeds, and a narrower boundary (0.2%~0.5% of image width) is used at lower speeds. At a speed of 5m / s, which is in the middle, a cropping range of 1.0% of the image width is used. For the pixel centroid coordinates, since the roll and pitch angles both have slight variations, the adjustment range is determined by multiplying the attitude stability coefficient by the maximum allowable range (assuming the maximum allowable adjustment is 5 pixels, then 0.533×5=2.665 pixels), and an approximate value of 3 pixels is used for offset adjustment. The specific values ​​of the final geometric adjustment parameter set are shown in Table 2.

[0078] Table 2. Set of Geometric Adjustment Parameters:

[0079]

[0080] As shown in Table 2, parameter adjustments ensure accurate execution of the adjustment operations based on the data, meeting the data requirements of the actual application environment. The formula determines the attitude stability coefficient through comprehensive calculation of normalized attitude, velocity, and acceleration parameters, making the basis for geometric parameter adjustments more specific and quantitative.

[0081] Please see Figure 3 The specific steps for obtaining the illumination compensation parameter set are as follows:

[0082] S211: Call the geometric adjustment parameter set, calculate the solar illumination direction vector and the shooting field of view direction vector based on the shooting timestamp and image acquisition location data, obtain the angle change trend of each time period, and generate the illumination angle change sequence.

[0083] The system calls upon a set of geometric adjustment parameters (rotation angle of 0.335°, boundary cropping range of 1% of image width, and pixel centroid offset of 3 pixels). Based on the shooting timestamps recorded by the UAV aerial photography system (e.g., the times of three consecutive sampling points are 08:00:00, 08:00:10, and 08:00:20) and image acquisition location data (e.g., latitude 30.5°N, longitude 114.3°E, altitude 150m), the UAV's built-in inertial navigation and positioning system calculates the solar altitude and azimuth angles at each time point in real time (e.g., solar altitude angle of 30.5° and azimuth angle of 120.5° at 08:00:00). This information is then converted into a unit vector and denoted as the solar illumination direction vector. Based on the drone's field-of-view orientation information (pitch angle 2.5°, roll angle -1.8°, heading angle 45.6°), the corresponding unit vector is calculated and denoted as the field-of-view orientation vector. Two unit vectors are multiplied by their inner product ( The included angle is calculated. Based on continuous time points, the included angle from 08:00:00 to 08:00:10 is 25°, and the included angle from 08:00:10 to 08:00:20 is 27°. Based on this, the trend of the included angle change is determined point by point, and finally a complete sequence of light angle change is obtained.

[0084] S212: Based on the sequence of changes in the illumination angle, analyze the shift trend of the solar illumination direction and the shooting field of view over a continuous time series, collect the corresponding illumination stability sequence and illumination change shift sequence, and use the formula:

[0085] ;

[0086] Calculate the exposure balance metric coefficient, evaluate the balance of exposure conditions, and establish exposure status labels;

[0087] in, For the first The normalized component of the sun's illumination direction at any given time is calculated by combining the shooting timestamp and geographical location with the solar altitude angle. For the first The normalized component of the field of view is captured at all times and calculated using the camera's current orientation and lens angle. For the first The normalized component of real-time lighting stability is obtained by normalizing a continuous sequence of light intensity collected by an ambient light intensity sensor. For the first The absolute value of the illumination change offset at any given time is calculated by the change in the average grayscale value of consecutive frames. The number of time-series sampling points is determined by dividing the sampling interval into aerial photography time windows. This serves as an ambient light reference constant, determined by collecting data on illumination variations under typical scenarios. This is the index number of the current sampling point in the time series, obtained by sequentially numbering the data at each moment within the shooting time window. This is a metric for exposure balance.

[0088] Based on the sequence of changes in illumination angle, the change in the angle between the direction of sunlight and the direction of the shooting field of view at each sampling time point was first quantitatively evaluated. The light intensity value at each time point was obtained by monitoring and collecting the ambient light intensity sensor, as shown in Table 3.

[0089] Table 3. Ambient Lighting and Image Data Acquisition:

[0090]

[0091] The ambient light intensity values ​​were normalized, with the highest light intensity of 32000 lx as the benchmark, and the normalized components of the ambient light intensity were obtained for each sampling point. The calculation process is as follows: the first point The second point The third point Component of the direction of solar illumination and field of view direction components The data in the table is used directly without normalization (it is already a unit vector component), and the absolute difference at point 1 is... Point 2 Point 3 Simultaneously calculate the absolute value of the image grayscale change. For example, the grayscale difference between points 1 and 2 The grayscale difference between the second and third points Take the ambient light reference constant The value is 20, which was determined experimentally within a typical range of ambient light variations (10–30).

[0092] Substitute the parameters into the exposure equalization metric formula for calculation:

[0093] ;

[0094] The specific calculation process is as follows:

[0095] ;

[0096] ;

[0097] ;

[0098] The final exposure balance metric coefficient is obtained as follows:

[0099] ;

[0100] The exposure balance metric coefficient is a dimensionless parameter that quantitatively reflects the consistency between the direction of sunlight and the shooting direction, the stability of lighting, and the overall amplitude of lighting changes during aerial photography. It directly measures the overall stability of the local and global lighting environment during image acquisition, as well as the coordination between the light source and camera direction. A lower value indicates that the light source and framing are coordinated, the lighting is uniform, and the lighting fluctuations are small, resulting in an ideal exposure state, suitable for using default exposure or slight adjustments. A higher value means that the lighting direction deviates significantly or the lighting is extremely unstable, easily leading to underexposure or overexposure, requiring active adjustment of exposure parameters and image compensation schemes. The coefficient serves as a dynamic criterion for the exposure state during drone photography, used for intelligent adjustment of preprocessing parameters, and is a key quantitative basis for achieving fully automatic lighting balance optimization. This result indicates a balanced exposure state; a value close to 0 indicates minimal exposure variation, therefore the exposure state label is determined as balanced exposure.

[0101] S213: Based on the exposure status label and the current scene's exposure status, adjust the image preprocessing configuration, including exposure time adjustment level, sensitivity control coefficient, image brightness correction intensity, and contrast enhancement ratio parameters, and generate a set of illumination compensation parameters.

[0102] Call the exposure status label (exposure status equalization), and adjust the image preprocessing configuration according to the exposure status. The specific operation is as follows: If the exposure equalization coefficient... If the image is located in the [0.0, 0.2] range, which is considered a balanced range, then the exposure time adjustment level should be set to a low level (e.g., level 1), with a short adjustment time (e.g., 2ms); the ISO control factor should be set to normal (ISO 100~200), taking the median value of ISO 150; the image brightness correction intensity should be determined by the difference between the exposure factor and the preset standard threshold (the threshold is set to 0.1), and the difference should be... The correction intensity is determined by multiplying the difference by a coefficient of 100, which is 3.19%, approximately 3.2%; the contrast enhancement ratio is set to a low ratio of 1.05 times in the balanced state; the specific values ​​of the illumination compensation parameter set composed of the above parameters are as follows: exposure time level is 1 (2ms), ISO is 150, brightness correction intensity is 3.2%, and contrast enhancement ratio is 1.05 times.

[0103] Please see Figure 4 The specific steps for obtaining the enhanced adjustment configuration are as follows:

[0104] S311: Call the illumination compensation parameter set, calculate the brightness change frequency, edge texture jitter and grayscale distribution fluctuation rate of the image block region, form the noise intensity index of each partition, and construct a noise index set;

[0105] Using the aforementioned lighting compensation parameter set (exposure time level 1, i.e., 2ms, ISO sensitivity set to 150, brightness correction intensity 3.2%, contrast enhancement ratio 1.05x), a single aerial image frame is divided into 9 partitions (3×3 grid), each 100×100 pixels. The brightness change frequency, edge texture jitter, and grayscale distribution fluctuation rate of each partition are calculated. The brightness change frequency is calculated by counting the proportion of pixels within the region whose grayscale difference between adjacent pixels exceeds a threshold (the threshold is set to 10 grayscale levels). Taking the first partition as an example, if there are 250 pixels whose grayscale difference between adjacent pixels exceeds the threshold, then the brightness change frequency is 250 / (100×100) = 0. .025; Edge texture jitter is obtained by averaging the absolute value of the grayscale changes of edge pixels in the region after edge detection. For example, if there are 1000 edge pixels in the first partition after edge detection, and the total grayscale change is 3000, then the edge texture jitter is 3000 / 1000 = 3; Grayscale distribution volatility is obtained by statistically analyzing the deviation of pixel grayscale values ​​from the average grayscale value of the region. Assuming the average grayscale of the first partition is 128, and the average difference between the grayscale values ​​of all pixels and the average grayscale is 12, then the grayscale distribution volatility is 12; The above indicators are calculated by normalizing and weighting (giving a weight of 0.3 for brightness change frequency, 0.4 for edge texture jitter, and 0.3 for grayscale distribution volatility), specifically, the noise intensity of the first partition is... The same process is used to calculate other partitions to obtain a set of noise metrics.

[0106] S312: Based on the noise index set, average the indices for each region of the image to form a full-image reference index set. Compare the indices of each region with the reference indices using the following formula:

[0107] ;

[0108] Calculate the overall noise shift magnitude, evaluate the noise shift magnitude of each region, and construct a region noise proportion matrix;

[0109] in, For the first The noise intensity of a partition is obtained by normalizing and weighting the frequency of brightness changes, edge texture jitter, and grayscale distribution fluctuation rate of that partition. The weighted average of noise intensity across the entire image is obtained by taking the weighted average of noise intensity across all regions. The root mean square of the noise intensity across the entire image is calculated by taking the root mean square of the noise intensity in all regions. For the first The spatial variation rate of noise intensity in a zone is calculated by normalizing the difference in noise intensity between adjacent zones. This is a normalization smoothing constant, a positive real number, obtained by looking up a preset parameter table. For the first Zonal composite noise offset magnitude For partitioned indexes;

[0110] Based on the noise intensity sets of the nine partitions obtained, as shown in Table 4:

[0111] Table 4. Noise intensity data for each zone:

[0112]

[0113] Calculate the weighted average noise intensity of the entire image based on the set of noise indicators. That is, the average value of the noise intensity of each zone:

[0114] ;

[0115] Then calculate the root mean square of noise intensity for the entire image. :

[0116] ;

[0117] Set normalization smoothing constant The value is 1.0, derived empirically from a preset parameter table, and ranges from 1 to 2; the spatial change rate is calculated for partition 1. The spatial variation rate is calculated by the normalized absolute value of the noise intensity difference between the first and adjacent zones. For example, if the average absolute value of the noise difference between the first and adjacent zones 2, 4, and 5 is 0.36, then the spatial variation rate is 0.36. Substituting the parameters into the formula:

[0118] ;

[0119] Taking partition 1 as an example, the specific calculation is as follows:

[0120] ;

[0121] The overall noise offset magnitude is a dimensionless parameter reflecting the relative degree of abnormality in the noise characteristics of each region in the image. It considers not only the offset of the noise intensity of the region from the mean of the entire image, but also the difference between the root mean square of the region and the entire image, as well as the spatial fluctuation of the noise intensity in the region. The larger the value, the more serious the deviation of the noise level of the region from the overall state, or the drastic noise change in the region, requiring more denoising or enhancement measures. The smaller the value, the more consistent the noise performance of the region is with the whole, and no excessive processing is required. The parameter provides an intuitive quantitative basis for subsequent adaptive image denoising and enhancement strategies, realizing fine-grained regional processing based on spatial heterogeneity, and effectively improving the image quality balance and detail expression ability. The same process is used to calculate the overall noise offset magnitude of other regions, forming a regional noise ratio matrix.

[0122] S313: Based on the regional noise ratio matrix, adjust the noise filtering intensity of each region according to the noise offset amplitude of each region to establish an enhancement adjustment configuration;

[0123] Based on the regional noise ratio matrix and the comprehensive noise offset amplitude of each partition, the comprehensive noise offset amplitude of partition 1 is 0.0708. This value is compared with the preset noise offset amplitude threshold range (0-0.05 for low noise offset, 0.05-0.1 for medium noise offset, and above 0.1 for high noise offset). Partition 1 is determined to be in the medium offset range, so the noise filtering intensity is moderate. The filtering intensity is adjusted by multiplying the basic filtering parameters (e.g., Gaussian filter radius of 1.0 pixel) by the offset amplitude ratio to 1.0 × (1 + 0.0708) = 1.0708 pixels. Other partitions are calculated in the same way. If the offset amplitude is high (e.g., exceeding 0.1), the filtering radius is increased to the range of 1.1-1.2 pixels. If the offset is low (less than 0.05), the filtering radius is close to the basic filtering radius of 1.0 pixel. After adjustment, an enhancement adjustment configuration is formed, and the specific filtering intensity of each region is clearly quantified for use in the actual image processing process.

[0124] Please see Figure 5 The specific steps for obtaining the scene configuration matching results are as follows:

[0125] S411: Using enhanced adjustment configuration, analyze the color distribution discreteness, boundary direction aggregation and texture concentration of image block regions, construct the semantic density vector of each partition, and generate a partition semantic density set;

[0126] Using enhanced adjustment configurations (e.g., a region filter intensity of 1.0708 pixels), the image is divided into 3×3 grid blocks, each region being 100×100 pixels. By invoking the specific action calculation process for color distribution dispersion, the mean and variance of the color distribution are calculated based on the RGB pixel values ​​of each region. For example, in the first region, the red channel has an average of 120 and a variance of 15, the green channel has an average of 125 and a variance of 12, and the blue channel has an average of 130 and a variance of 10. The variances of the three channels are then summed using weights (red 0.4, green 0.3, blue 0.3) to obtain the color distribution dispersion. The boundary direction aggregation is calculated using edge gradient direction statistics, with the Sobel operator performing gradient calculations in 8 directions. For example, the first partition has the highest proportion of pixels along the horizontal direction, and its aggregation value is calibrated to 0.65 (range 0~1, the closer to 1, the stronger the aggregation). Texture concentration is calculated using the gray-level co-occurrence matrix entropy value; the smaller the entropy value, the stronger the texture concentration. For example, the texture entropy value of the first partition is 3.8, and the normalized concentration index is... The three indicators are combined to form the semantic density vector of the first partition (12.9, 0.65, 0.525). All partitions are calculated in this way to form a set of partition semantic densities.

[0127] S412: Based on the semantic density set of the partition, compare the coverage and proportion trend of each semantic cluster category in the image, determine the category dominance and identify the scene label of each frame image, and obtain the scene label discrimination result;

[0128] Based on the semantic density set of each partition, the vector dissimilarity of each semantic category in each partition is calculated, specifically using Euclidean distance. For example, the Euclidean distance between the vector (12.9, 0.65, 0.525) of partition 1 and the vector (10, 0.7, 0.6) of the standard semantic category "road" is:

[0129] ;

[0130] The distance to the "Architecture" category vector (14, 0.5, 0.4) is then calculated as follows: ;

[0131] The distances to different categories for each region were calculated separately, and the category with the smallest distance was selected as the dominant category for that region. The category for the first sub-region was "Building". The distribution data of the dominant category for all sub-regions is shown in Table 5.

[0132] Table 5: Statistics of Dominant Categories in Each Region

[0133]

[0134] Based on the statistical data in Table 5, the coverage rate of each category was calculated. The coverage rate of "buildings" was 4 / 9≈44.4%, "roads" was 3 / 9≈33.3%, and "green space" was 2 / 9≈22.2%. By comparing the coverage rate and proportion trend of each category, the "buildings" category had the highest coverage rate and was determined to be the dominant category of the image. The scene label of this frame image was determined to be "building-dominated scene", and the scene label discrimination result was obtained.

[0135] S413: Based on the scene label discrimination result, match the corresponding image processing parameters according to the scene label of the image to obtain the scene configuration matching result;

[0136] Based on the determined "architectural dominant scene" scene label, a preset image processing parameter library is matched. This library contains scene labels and corresponding parameter matching settings, as detailed in Table 6.

[0137] Table 6: Scenario Configuration Parameter Matching Table

[0138]

[0139] As shown in Table 6, the "Architecture-Dominated Scene" tag is matched with a sharpening intensity of 15%, a saturation adjustment coefficient of 1.05, a contrast adjustment coefficient of 1.10, and a noise threshold of 0.03. The scene configuration matching result is generated by calling the action through the parameters.

[0140] Please see Figure 6 The specific steps for obtaining optimized aerial image records are as follows:

[0141] S511: Using the scene configuration matching results and the flight altitude corresponding to each image frame, calculate the ground projection line-of-sight trajectory of each pixel in the image under the current viewpoint, and obtain the pixel projection line-of-sight set.

[0142] Using the sharpening intensity (15%), saturation adjustment coefficient (1.05), contrast adjustment coefficient (1.10), and noise threshold (0.03) obtained from the scene configuration matching results, combined with the flight altitude corresponding to the image frames recorded by the UAV flight control (e.g., 150 meters), the camera lens angle parameters (lens pitch angle 2.5°, lens roll angle -1.8°) of each pixel in the image are called one by one. By actually calculating the projection distance of each pixel relative to the ground, taking the image center pixel as the origin (0,0), and its horizontal and vertical pixel spacing as a reference of 0.01 meters and 0.01 meters respectively, for the pixel located at the offset position (100,150) of the center point (0,0), the projection distance is calculated using the similar triangle relationship: first, the horizontal offset distance is obtained by multiplying the pixel offset by the pixel physical spacing. meters) and vertical offset distance ( Then, based on the relationship between the flight altitude of 150 meters and the camera's pitch and roll attitude angles (pitch 2.5°, roll -1.8°), the actual coordinates of the pixel's projection on the ground are obtained through calculation using the angle cosine relationship. The corresponding ground projection line of sight is the actual ground distance from the UAV's vertical projection point (for example, the calculated actual ground projection coordinates of this pixel are (1.2 meters, 1.7 meters)). Subsequently, the actual projection line of sight of the pixel is calculated using the Pythagorean theorem. Meters are used to calculate the total pixel distance of the image, one by one, to form a complete set of pixel projection distances.

[0143] S512: Based on the pixel projection viewing distance set, analyze the projection change span between pixels, form a spatial gradient distribution, construct a two-dimensional spatial gradient change heat map, and obtain spatial gradient heat distribution information.

[0144] Based on the pixel projection viewing distance set, the projection viewing distance between adjacent pixels is calculated and compared point by point. Specifically, the projection change span is analyzed by calculating the distance difference between adjacent pixels. Taking two adjacent pixels as an example, if one pixel's ground projection viewing distance is 2.08 meters and the other adjacent pixel's ground projection viewing distance is calculated to be 2.20 meters, then the distance change span between them is calculated as follows: The algorithm iterates through the entire image region pixel by pixel, performing the same calculation on all adjacent pixels. All calculated distance variation values ​​are then filled into a two-dimensional matrix of the same size as the original image. Each matrix element records the distance variation value between the corresponding pixel and its surrounding pixels. The matrix values ​​are then standardized by dividing all span values ​​by the maximum span of the entire image (e.g., a maximum span of 0.50 meters), thus normalizing the span values ​​and forming a spatial gradient distribution with values ​​ranging from [0,1]. Next, all normalized span values ​​are used to convert the gradient value at each pixel location into a color map. Normalized gradient values ​​close to 1 are set as red areas, indicating a large projection span, while values ​​close to 0 are set as blue areas, indicating a small projection span. This forms a complete two-dimensional spatial gradient change heatmap, thereby obtaining spatial gradient heat distribution information.

[0145] S513: Based on the spatial gradient thermal distribution information, the pixel region is divided into multiple levels according to the degree of continuous span, and the texture reconstruction intensity parameters of each level are adjusted to obtain the aerial image optimization record.

[0146] Based on the acquired spatial gradient thermal distribution information, the pixels of the entire image region are classified into different levels according to the gradient normalization value. Normalized gradient values ​​in the range of 0-0.3 are defined as "low gradient regions," 0.3-0.6 as "medium gradient regions," and 0.6-1.0 as "high gradient regions." This is determined by iterating through the gradient values ​​of each pixel. For example, a pixel with a normalized gradient value of 0.72 belongs to the high gradient region level. After the pixel level is determined, the texture reconstruction intensity parameters corresponding to each level are called one by one. Based on a preset basic texture reconstruction intensity of 1.0 (with a reasonable range of 0.8-1.2 determined experimentally), the parameters are adjusted according to the different gradient levels of the regions. The texture reconstruction intensity for low gradient regions is set to 0.8 (reduced reconstruction), for medium gradient regions to 1.0 (basic reconstruction), and for high gradient regions to 1.2 (enhanced reconstruction). The texture reconstruction intensity is determined for each pixel according to its corresponding level, and the parameters are applied to complete the texture optimization operation. The optimized texture reconstruction intensity configuration for each pixel region is recorded, forming a detailed aerial image optimization record.

[0147] An AI-based UAV aerial image optimization system, comprising: [details omitted for brevity]

[0148] The attitude trajectory analysis module calls upon flight status information to analyze the continuous change trajectory of the UAV's pitch angle, roll angle, and heading angle in the time series, assesses the degree of attitude deviation of the UAV, and obtains a set of geometric adjustment parameters based on the direction and speed information of angle change.

[0149] The exposure equalization evaluation module calls the geometric adjustment parameter set, calls the shooting timestamp, calculates the angle range between the sun's illumination direction and the shooting field of view direction when the image is acquired, evaluates the equalization state of the exposure conditions, adjusts the image preprocessing configuration, and obtains the illumination compensation parameter set.

[0150] The regional noise quantization module calls the illumination compensation parameter set to calculate the brightness change frequency, edge texture jitter, and grayscale distribution fluctuation rate in the image block regions, constructs a noise index set, averages the index of each region to form a full-image reference index set and compares it with each region, evaluates the noise offset amplitude of each region and adjusts the noise filtering intensity to obtain the enhancement adjustment configuration.

[0151] The image semantic clustering module uses enhanced adjustment configuration to analyze the color distribution discreteness, boundary direction aggregation and texture concentration in image block regions, constructs semantic density vectors, identifies scene labels and configures image processing parameters, and obtains scene configuration matching results.

[0152] The spatial gradient reconstruction module uses scene configuration matching results and flight altitude information to calculate the spatial gradient changes formed by pixel projection line of sight, constructs a gradient change heatmap in two-dimensional space, divides the pixel region into multiple levels according to the degree of continuous span in the gradient map, adjusts the texture reconstruction intensity parameters, and obtains optimized records of aerial images.

[0153] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for optimizing drone aerial images based on artificial intelligence, characterized in that, Includes the following steps: S1: Call the navigation status information, analyze the continuous change trajectory of the UAV's pitch angle, roll angle and heading angle in the time series, assess the degree of attitude deviation of the UAV, and obtain the set of geometric adjustment parameters based on the angle change direction and speed information; The specific steps for obtaining the geometric adjustment parameter set are as follows: S111: Call the navigation status information to obtain continuous time series data of the UAV's pitch angle, roll angle and heading angle, analyze the angle change direction and velocity change status of each sampling time period, and establish an angle change dataset; S112: Based on the angle change dataset, calculate the range of change of each angle in a continuous time period, normalize the change of each angle, calculate the attitude stability coupling coefficient, and output the attitude offset dynamic index. S113: Based on the attitude offset dynamic index, combined with the angle change direction and flight speed trend, configure the adjustment parameters of image rotation angle, boundary cropping range and pixel centroid coordinates to generate a geometric adjustment parameter set; S2: Call the geometric adjustment parameter set, call the shooting timestamp, calculate the angle range between the sun's illumination direction and the shooting field of view during image acquisition, evaluate the balance of exposure conditions, adjust the image preprocessing configuration, and obtain the illumination compensation parameter set; S3: Call the illumination compensation parameter set, calculate the brightness change frequency, edge texture jitter and gray scale distribution fluctuation rate in the image block area, construct a noise index set, average the index of each region to form a full-image reference index set and compare it with each region, evaluate the noise offset amplitude of each region and adjust the noise filtering intensity to obtain the enhancement adjustment configuration; S4: Using the enhanced adjustment configuration, analyze the color distribution discreteness, boundary direction aggregation and texture concentration in the image block region, construct a semantic density vector, identify scene labels and configure image processing parameters, and obtain scene configuration matching results; The specific steps for obtaining the scenario configuration matching result are as follows: S411: Using the aforementioned enhanced adjustment configuration, analyze the color distribution dispersion, boundary direction aggregation, and texture concentration of the image block region. The boundary direction aggregation is calculated using edge gradient direction statistics, with the Sobel operator performing gradient calculations in 8 directions. The three indicators are combined to construct the semantic density vector of each partition, generating a set of partition semantic densities. S412: Based on the partition semantic density set, compare the coverage and proportion trend of each semantic clustering category in the image, determine the category dominance and identify the scene label of each frame image, and obtain the scene label discrimination result; S413: Based on the scene label discrimination result, match the corresponding image processing parameters according to the scene label of the image to obtain the scene configuration matching result.

2. The method for optimizing UAV aerial images based on artificial intelligence according to claim 1, characterized in that, The geometric adjustment parameter set specifically includes rotation angle configuration data, cropping boundary coordinate information, and repositioning pixel center data. The illumination compensation parameter set includes exposure control level, photosensitivity response coefficient, brightness adjustment parameters, and contrast enhancement factor. The enhancement adjustment configuration specifically includes filter response adjustment level, regional noise suppression ratio, and image enhancement control factor. The scene configuration matching result includes image spatial label set, regional semantic recognition number, and processing template selection mark.

3. The method for optimizing UAV aerial images based on artificial intelligence according to claim 1, characterized in that, The specific steps for obtaining the illumination compensation parameter set are as follows: S211: Call the geometric adjustment parameter set, calculate the solar illumination direction vector and the shooting field of view direction vector based on the shooting timestamp and image acquisition location data, obtain the angle change trend of each time period, and generate the illumination angle change sequence. S212: Based on the light angle change sequence, analyze the offset trend of the sun's illumination direction and the shooting field of view in a continuous time series, collect the corresponding lighting stability sequence and light change offset sequence, calculate the exposure balance measurement coefficient, evaluate the balance state of the exposure conditions, and establish an exposure state label. S213: Based on the exposure status label, adjust the image preprocessing configuration according to the current scene's exposure status, including exposure time adjustment level, sensitivity control coefficient, image brightness correction intensity, and contrast enhancement ratio parameters, and generate a set of illumination compensation parameters.

4. The method for optimizing UAV aerial images based on artificial intelligence according to claim 3, characterized in that, The specific steps for obtaining the enhanced adjustment configuration are as follows: S311: Call the illumination compensation parameter set to calculate the brightness change frequency, edge texture jitter and grayscale distribution fluctuation rate of the image block region, form the noise intensity index of each partition, and construct a noise index set; S312: Based on the noise index set, the indices of each region of the image are averaged to form a full-image reference index set. By comparing the indices of each region with the reference indices, the comprehensive noise offset amplitude is calculated, the noise offset amplitude of each region is evaluated, and a region noise ratio matrix is ​​constructed. S313: Based on the aforementioned regional noise ratio matrix, adjust the noise filtering intensity of each region according to the noise offset amplitude of each region to establish an enhancement adjustment configuration.

5. The method for optimizing UAV aerial images based on artificial intelligence according to claim 1, characterized in that, The method further includes: S5: Using the scene configuration matching results and flight altitude information, calculate the spatial gradient change formed by pixel projection line of sight, construct a gradient change heat map in two-dimensional space, divide the pixel region into multiple levels according to the degree of continuous span in the gradient map, adjust the texture reconstruction intensity parameters, and obtain the aerial image optimization record. The aerial image optimization record includes depth partition mapping identifiers, region reconstruction parameter levels, and texture processing control groups.

6. The artificial intelligence-based drone aerial image optimization method according to claim 5, characterized in that, The specific steps for obtaining the optimized aerial image record are as follows: S511: Using the scene configuration matching results and the flight altitude corresponding to each image frame, calculate the ground projection line-of-sight trajectory of each pixel in the image under the current viewpoint, and obtain the pixel projection line-of-sight set. S512: Based on the pixel projection viewing distance set, analyze the projection change span between pixels, form a spatial gradient distribution, construct a two-dimensional spatial gradient change heat map, and obtain spatial gradient heat distribution information. S513: Based on the spatial gradient thermal distribution information, the pixel region is divided into multiple levels according to the degree of continuous span, and the texture reconstruction intensity parameters of each level are adjusted to obtain the aerial image optimization record.

7. An artificial intelligence-based drone aerial image optimization system, characterized in that, The system is used to implement the artificial intelligence-based drone aerial image optimization method according to any one of claims 1-6, and the system comprises: The attitude trajectory analysis module calls upon flight status information to analyze the continuous change trajectory of the UAV's pitch angle, roll angle, and heading angle in the time series, assesses the degree of attitude deviation of the UAV, and obtains a set of geometric adjustment parameters based on the direction and speed information of angle change. The exposure equalization evaluation module calls the geometric adjustment parameter set, calls the shooting timestamp, calculates the angle range between the sun's illumination direction and the shooting field of view direction when the image is acquired, evaluates the equalization state of the exposure conditions, adjusts the image preprocessing configuration, and obtains the illumination compensation parameter set. The regional noise quantization module calls the illumination compensation parameter set to calculate the brightness change frequency, edge texture jitter and grayscale distribution fluctuation rate in the image block region, constructs a noise index set, averages the index of each region to form a full-image reference index set and compares it with each region, evaluates the noise offset amplitude of each region and adjusts the noise filtering intensity to obtain the enhancement adjustment configuration. The image semantic clustering module uses the enhanced adjustment configuration to analyze the color distribution discreteness, boundary direction aggregation and texture concentration in the image block region, constructs a semantic density vector, identifies scene labels and configures image processing parameters, and obtains scene configuration matching results. The spatial gradient reconstruction module uses the scene configuration matching results and flight altitude information to calculate the spatial gradient change formed by the pixel projection line of sight, constructs a gradient change heatmap in two-dimensional space, divides the pixel region into multiple levels according to the degree of continuous span in the gradient map, adjusts the texture reconstruction intensity parameters, and obtains an optimized record of aerial images.

Images