High-definition image processing and transmission method based on unmanned aerial vehicle

Abstract

The invention relates to the technical field of unmanned aerial vehicle image processing, and discloses a high-definition image processing and transmission method based on an unmanned aerial vehicle. The method comprises the following steps: acquiring a multi-frame high-definition image sequence obtained in the flight process of an unmanned aerial vehicle; analyzing the image sequence frame by frame, extracting key feature points and generating feature descriptors; establishing a feature matching relationship between adjacent frames, and calculating an inter-frame motion vector; constructing a global motion model according to the motion vector and performing motion compensation; performing region segmentation on the compensated image sequence to divide the image sequence into a dynamic region and a static region; performing multi-scale noise reduction on the dynamic region, and performing texture enhancement on the static region; carrying out fusion reconstruction on the processed dynamic and static regions to generate an optimized image sequence; performing layered coding on the optimized sequence according to a transmission bandwidth condition, and generating base layer and enhancement layer data; dynamically adjusting the priority of two layers of data transmission according to the network state; and sending the data to a ground receiving end through a multi-path transmission protocol.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle image processing, in particular to a high-definition image processing and transmission method based on unmanned aerial vehicles. BACKGROUND

[0002] In recent years, unmanned aerial vehicle technology has developed rapidly. The high-definition camera equipment carried by the unmanned aerial vehicle can capture a large amount of continuous image sequence data during flight. These data have wide application value in many fields such as environmental monitoring, topographic mapping, emergency rescue, and security patrol. However, due to the instability of the unmanned aerial vehicle platform and environmental factors such as air flow disturbance and flight attitude change, the captured image sequence will have different degrees of shaking and blurring. This motion blur will seriously affect the accuracy of subsequent image content recognition and analysis. In addition, the working environment of the unmanned aerial vehicle is complex and changeable, and its image data usually contains static background scenes and dynamic moving targets, such as moving vehicles or moving personnel. It is often difficult to achieve ideal results by using a unified image processing strategy for these two types of areas. The static background may need to enhance the detail clarity, while the dynamic target needs to suppress noise and motion blur.

[0003] In terms of data transmission, the wireless communication link between the unmanned aerial vehicle and the ground station has limited and unstable bandwidth, especially at long distances or in complex electromagnetic environments. Direct transmission of high-definition image sequences without processing not only occupies a large amount of channel resources, causing transmission delay to increase, but also may cause data packet loss due to network congestion, affecting the integrity and real-time performance of image reconstruction at the ground receiving end. Although some existing image transmission schemes attempt to reduce data volume through compression encoding, most lack targeted optimization of the characteristics of the image content itself, such as failing to effectively distinguish between dynamic and static regions in the image. At the same time, traditional transmission strategies usually use fixed encoding and transmission modes, which are difficult to adaptively respond to dynamic changes in network state, resulting in a sharp decline in image quality when bandwidth is tight, or failing to fully utilize resources to improve image quality when bandwidth is abundant. Therefore, there is an urgent need for a method that can optimize image processing and data transmission together, improving the visual quality of high-definition image sequences from unmanned aerial vehicles while ensuring efficient and reliable transmission under unstable network conditions. SUMMARY

[0004] The present application aims to provide a high-definition image processing and transmission method based on unmanned aerial vehicles to solve the problems raised in the background.

[0005] To achieve the above-mentioned purpose, the present application provides a high-definition image processing and transmission method based on unmanned aerial vehicles, which comprises: Collect a plurality of high-definition image sequences acquired in a flight process of a UAV, wherein the high-definition image sequences contain image data at different time stamps; Perform frame-by-frame analysis on the high-definition image sequences, extract key feature points in each frame of image, and generate image feature descriptors according to the key feature points; Establish a feature matching relationship between adjacent frames based on the image feature descriptors, and calculate inter-frame motion vectors; Construct a global motion model according to the inter-frame motion vectors, and perform motion compensation on the high-definition image sequences through the global motion model; Perform region segmentation on the high-definition image sequences after motion compensation, and divide dynamic regions and static regions; Perform multi-scale noise reduction processing on the dynamic regions, and perform texture enhancement processing on the static regions; Fuse and reconstruct the processed dynamic regions and static regions to generate an optimized high-definition image sequence; According to a preset transmission bandwidth condition, perform layered encoding on the optimized high-definition image sequence to generate base layer data and enhancement layer data; Based on network status, dynamically adjust the transmission priorities of the base layer data and the enhancement layer data; Send the adjusted base layer data and enhancement layer data to a ground receiving end through a multi-path transmission protocol.

[0006] Preferably, the extraction of key feature points in each frame of image comprises: Identify contour information in the image using an edge detection algorithm, and filter out feature points with significant gradients according to the contour information; Perform local descriptor calculation on the feature points to generate feature vectors with rotation invariance; Cluster the feature points according to the similarity of the feature vectors, and remove redundant feature points.

[0007] Preferably, the establishment of a feature matching relationship between adjacent frames based on the image feature descriptors comprises: Calculate the Euclidean distance between the feature descriptors of the current frame and the previous frame to determine the initial matching pairs; Filter the initial matching pairs through a random sample consensus algorithm to remove mis-matching points; Calculate the inter-frame motion vectors according to the filtered matching points.

[0008] Preferably, the construction of a global motion model according to the inter-frame motion vectors comprises: Fit an affine transformation matrix using the inter-frame motion vectors to describe the inter-frame global motion trend; The parameter errors of the affine transformation matrix are corrected through an iterative optimization algorithm; Motion compensation is performed on subsequent frames using a modified affine transformation matrix.

[0009] Preferably, the region segmentation of the motion-compensated high-definition image sequence includes: Calculate the inter-frame difference map for the motion-compensated image sequence and mark the regions where pixel changes exceed a threshold; Based on the optical flow field analysis results, the dynamic and static regions are distinguished. Morphological processing is performed on the dynamic region to eliminate erroneous segmentation caused by noise.

[0010] Preferably, the multi-scale noise reduction processing of the dynamic region includes: Wavelet decomposition is performed on the dynamic region to obtain sub-images of different frequency bands; Threshold filtering is applied to high-frequency sub-images based on noise statistical characteristics; Reconstruct the filtered sub-image to preserve the details of the dynamic regions.

[0011] Preferably, the simultaneous texture enhancement processing of the static region includes: Local binary mode analysis is performed on the static region to extract texture features. The contrast enhancement coefficient is adaptively adjusted based on the texture features, and a nonlinear filtering algorithm is used to enhance the edge structure of the static region.

[0012] Preferably, the step of performing layered encoding on the optimized high-definition image sequence includes: The optimized high-definition image sequence is subjected to discrete cosine transform to separate low-frequency and high-frequency components; The low-frequency components are encoded as the base layer data, and the high-frequency components are encoded as the enhancement layer data; The base layer data and enhancement layer data are compressed using an entropy coding algorithm.

[0013] Preferably, the dynamic adjustment of the transmission priority of the base layer data and enhancement layer data based on network status includes: Real-time monitoring of network packet loss rate and latency, and calculation of the current channel quality index; The transmission bandwidth ratio of the base layer data and the enhancement layer data is dynamically allocated according to the channel quality index, and the base layer data is transmitted first when the network is congested.

[0014] Preferably, the step of sending the adjusted base layer data and enhancement layer data to the ground receiving end via a multipath transmission protocol includes: The base layer data and enhancement layer data are split into multiple data blocks, each data block is assigned an independent transmission path identifier, and a redundancy check mechanism is used to ensure the complete reception of the data blocks.

[0015] Compared with the prior art, the beneficial effects of the present invention are: The method of this invention offers multiple advantages through integrated image processing and adaptive transmission strategies. In the image preprocessing stage, it extracts key feature points and establishes an inter-frame motion model for precise motion compensation, effectively suppressing image blurring caused by UAV platform jitter and improving the overall stability and clarity of the sequence. Furthermore, it intelligently divides the image content into dynamic and static regions, applying multi-scale noise reduction and texture enhancement processing to each, resulting in clearer outlines of dynamic targets, effective noise suppression, and full preservation and enhancement of static background details, significantly improving overall image quality.

[0016] In the data transmission stage, a layered coding mechanism is employed to decompose the optimized image sequence into a base layer and an enhancement layer. The base layer carries the core information necessary for image reconstruction, while the enhancement layer contains supplementary details to improve image quality. This structure provides a high degree of flexibility in the transmission strategy. Combined with a multipath transmission protocol, the system can intelligently adjust the transmission priority of the two layers of data based on real-time network bandwidth conditions. For example, when network bandwidth is limited, priority is given to ensuring the complete transmission of the base layer data, ensuring that the ground station can receive usable basic image information; while when network conditions improve, the transmission of the enhancement layer data is gradually increased, thereby gradually improving the detail quality and visual effect of the received image. This dynamic adaptive mechanism effectively overcomes the limitations of traditional fixed-rate transmission in dealing with network fluctuations.

[0017] This method achieves a deep integration of image content optimization and transmission resource management. Targeted image processing at the front end creates favorable conditions for efficient compression and robust transmission at the back end, while the adaptive transmission strategy at the back end ensures that the front-end processing results can be delivered to users to the greatest extent possible in various network environments. This end-to-end optimization approach enables the UAV system to continuously provide higher-quality and more stable image data services in complex real-world application scenarios, enhancing the reliability and practicality of UAV operations. Ultimately, users can obtain high-definition images with better visual experience, richer details, and more reliable real-time performance, meeting the urgent needs of various professional applications for high-quality image data. Attached Figure Description

[0018] Figure 1 A comparison chart of peak signal-to-noise ratio and structural similarity at each stage of high-definition image processing for UAVs; Figure 2 A flowchart for extracting key feature points from each frame of an image; Figure 3 A flowchart for constructing a global motion model based on inter-frame motion vectors; Figure 4 This is a chart analyzing the proportion of hierarchically encoded data and the compression ratio. Detailed Implementation

[0019] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides a method for high-definition image processing and transmission based on unmanned aerial vehicles (UAVs). The method includes: acquiring a multi-frame high-definition image sequence during UAV flight, the high-definition image sequence containing image data at different timestamps; performing frame-by-frame analysis on the high-definition image sequence, extracting key feature points in each frame, and generating image feature descriptors based on the key feature points; establishing feature matching relationships between adjacent frames based on the image feature descriptors, and calculating inter-frame motion vectors; constructing a global motion model based on the inter-frame motion vectors, and performing motion compensation on the high-definition image sequence using the global motion model; performing region segmentation on the motion-compensated high-definition image sequence to divide it into dynamic and static regions; performing multi-scale noise reduction processing on the dynamic regions, and simultaneously performing texture enhancement processing on the static regions; fusing and reconstructing the processed dynamic and static regions to generate an optimized high-definition image sequence; performing layered encoding on the optimized high-definition image sequence according to preset transmission bandwidth conditions to generate base layer data and enhancement layer data; dynamically adjusting the transmission priority of the base layer data and enhancement layer data based on network status; and sending the adjusted base layer data and enhancement layer data to the ground receiving end through a multipath transmission protocol.

[0021] Example 1: See Figure 2Key feature points are extracted from each frame of the image, and an edge detection algorithm is used to identify contour information in the image. The Canny operator is selected for edge detection. The Canny operator processes the image data through a series of steps. In the initial stage, a Gaussian filter is applied to smooth the input high-resolution image sequence. The kernel size and standard deviation parameters of the Gaussian filter are configured according to the image resolution and noise level. The smoothed image reduces the interference of high-frequency noise on edge detection. The gradient magnitude and direction of each pixel in the image are calculated. The gradient calculation is completed by the Sobel operator or the Prewitt operator. The operator performs convolution operations with the image in the horizontal and vertical directions, respectively, to obtain the horizontal gradient component and the vertical gradient component. The gradient magnitude is the square root of the sum of the squares of the two components, and the gradient direction is determined by the arctangent of the two components. Non-maximum suppression is applied to the gradient magnitude matrix. Non-maximum suppression compares the gradient magnitude of the current pixel with the magnitudes of the two adjacent pixels along the gradient direction. If the magnitude of the current pixel is not the maximum value, its magnitude is set to zero. This process preserves the pixels corresponding to the local maximum values ​​of the gradient magnitude, thereby refining the width of the edges. In the dual-threshold detection stage, two thresholds are set: a high threshold and a low threshold. Pixels with gradient magnitudes higher than the high threshold are identified as strong edge points, while pixels with gradient magnitudes between the low and high thresholds are marked as weak edge points. Strong edge points are directly retained as final edges, while weak edge points must be connected to strong edge points to be retained as edges; otherwise, they are suppressed. The connection process is achieved by checking whether a strong edge point exists in the eight-neighborhood of a weak edge point. Feature points with significant gradients are selected based on contour information. Feature points with significant gradients are defined as pixels located on the final edge whose gradient magnitude exceeds the set high threshold. These points typically correspond to locations in the image where there are significant changes in geometric structure, such as corners and edge intersections of objects.

[0022] Local descriptor calculation is performed on feature points to generate rotation-invariant feature vectors. The local descriptor calculation employs a scale-invariant feature transform algorithm. This algorithm constructs descriptors in the image scale space where the feature points reside. The scale space is constructed using a Gaussian pyramid, which is generated by continuously applying Gaussian blur and downsampling to the original image to create multiple layers. The principal orientation of the feature points is determined by calculating the gradient orientation histogram of pixels in the feature point's neighborhood. This histogram divides the 360-degree direction into 36 pillars, each representing a 10-degree directional range. The gradient orientations of all pixels in the feature point's neighborhood are weighted according to their gradient magnitudes and then voted into the corresponding orientation pillar. The peak value of the histogram represents the principal orientation of the feature point. A feature descriptor vector is constructed by taking a neighborhood window of a certain size centered on the feature point and rotating the window to align with the main direction to ensure rotation invariance. The rotated window is then divided into 4x4 sub-regions. For each sub-region, the gradient direction histogram in its eight directions is calculated, resulting in an 8-dimensional vector for each sub-region. A total of 16 sub-regions generate a 128-dimensional feature vector. The feature vector is normalized to reduce the impact of illumination changes. Normalization divides each element of the vector by the magnitude of the entire vector, ensuring the feature vector remains stable under linear illumination changes. Feature points are clustered based on the similarity of the feature vectors, eliminating redundant feature points. The similarity of feature vectors is measured by Euclidean distance. The K-means clustering algorithm is applied to the 128-dimensional feature vector space of all feature points. The K-means clustering algorithm first randomly selects K feature vectors as initial cluster centers, with the value of K determined based on the image content complexity and the desired number of feature points to retain. Each feature vector is assigned to the nearest cluster center. The distance is calculated using the Euclidean distance formula, which is the square root of the sum of the squares of the differences in each dimension of the two vectors. The cluster center is updated to the mean of all feature vectors belonging to that cluster. This assignment and update process is repeated until the cluster centers no longer change significantly or the maximum number of iterations is reached. For each cluster generated after clustering, the feature points contained within the cluster are examined. Only the feature point with the largest gradient magnitude in the cluster is retained. The gradient magnitude is directly obtained from the gradient information saved during the feature point detection stage. Other feature points within the same cluster are considered different repetitive or approximate instances describing the same local structure in the image due to the similarity of their feature vectors. These points are marked as redundant feature points and removed from the feature point set. The feature point set after removing redundant feature points is more evenly distributed in the image space, and the number of feature points is significantly reduced, improving the computational efficiency of the subsequent feature matching stage.

[0023] Establishing feature matching relationships between adjacent frames based on image feature descriptors involves calculating the Euclidean distance between the feature descriptors of the current frame and the previous frame, and determining initial matching pairs. An initial matching pair refers to a pair of data formed by filtering a feature point in the current frame and a feature point in the previous frame using Euclidean distance; one matching pair contains two individual feature points. For each feature point in the current frame's feature point set, all feature points in the previous frame's feature point set are traversed, and the Euclidean distance between its corresponding 128-dimensional feature vector and the feature vector of the current frame's feature point is calculated. The feature point in the previous frame with the smallest Euclidean distance to the current frame's feature point is found, and this minimum distance value is recorded. A distance ratio threshold is set, and the distance between the current frame's feature point and the second smallest distance feature point in the previous frame is divided by the minimum distance. If this ratio is less than the set threshold, the current frame's feature point and the previous frame's feature point corresponding to the minimum distance are accepted as an initial matching pair. The distance ratio threshold is typically set between 0.6 and 0.8. This strategy utilizes the characteristic that the distance between a correctly matched point pair and its nearest neighbor should be significantly smaller than the distance between an incorrectly matched point pair. The initial matching pairs are filtered using a random sampling consensus algorithm (RSA), an iterative method for estimating model parameters, to remove mismatches. The RSA algorithm randomly selects four pairs of matching points from the initial set. These four pairs represent four independent sets of matching points, containing a total of eight individual feature points. These four pairs represent the minimum number of samples required to compute the affine transformation model. An initial affine transformation matrix model with six degrees of freedom is then calculated using the coordinate correspondence of the four pairs of matching points. The solved initial affine transformation matrix model is then used to test all other matching points in the initial set. The coordinates of the current frame feature points are projected onto the coordinate system of the previous frame using the affine transformation matrix. The Euclidean distance between the projected point and the corresponding feature point coordinates of the previous frame is calculated. If this distance is less than a set error tolerance threshold, the single matching point (i.e., the pair consisting of a single feature point in the current frame and its corresponding single feature point in the previous frame) is considered an interior point; otherwise, it is considered an exterior point. The system counts the number of inliers corresponding to the current model and records the model with the most inliers and its set of inliers. The random sampling, model calculation, and inlier counting process are repeated for a preset number of iterations. The number of iterations is determined by the desired model confidence and the expected proportion of inliers in the initial matching pairs. The model obtained from the iteration with the most inliers is selected as the final correct geometric transformation model, and the set of inliers corresponding to this model is taken as the filtered correct matching pairs. All initial matching pairs that are not classified as inliers are considered mismatches and are discarded.

[0024] Inter-frame motion vectors are calculated based on the selected matching point pairs. These vectors represent the global motion trend between adjacent frames. For the set of correctly matched point pairs obtained after filtering, the motion vector for each pair is calculated. The motion vector is a two-dimensional vector whose components are the x-coordinate of the current frame feature point minus the x-coordinate of the previous frame's matching feature point, and the y-coordinate of the current frame feature point minus the y-coordinate of the previous frame's matching feature point. Statistical processing is performed on the motion vectors of all correctly matched point pairs, calculating the median of the x and y components. The median is selected as the final estimate of the inter-frame motion vector. Using the median estimate instead of the mean effectively suppresses interference from possible local motion or remaining mismatched points on the global motion estimation. The resulting inter-frame motion vector is a two-dimensional vector used as input data for subsequently constructing the global motion model.

[0025] Example 2: See Figure 3 A global motion model is constructed based on inter-frame motion vectors. This construction involves fitting an affine transformation matrix using the inter-frame motion vectors. The affine transformation matrix is ​​a linear transformation model on a two-dimensional plane with six degrees of freedom, capable of describing translation, rotation, scaling, and shearing motions between images. The fitting process is based on a filtered set of correctly matched point pairs, selected using the random sampling consensus algorithm described in the previous embodiment. Each correctly matched point pair contains the coordinates of a feature point in the current frame and the coordinates of a corresponding feature point in the previous frame. The mathematical model of the affine transformation maps the coordinates of the current frame point to the coordinate system of the previous frame. This mapping relationship is represented by an equation containing six parameters, involving a linear combination of the horizontal and vertical coordinates of the current frame point and the parameters of the transformation matrix. The least squares method is used to solve for the optimal parameters of the affine transformation matrix. The least squares method constructs an error function, defined as the sum of the squared Euclidean distances between the projected points of all correctly matched point pairs after the affine transformation and their actual matching points. The least squares method obtains a normal equation system by solving the partial derivatives of the error function with respect to the six parameters and setting them to zero. Solving this linear equation system yields the six affine transformation parameters that minimize the sum of squared projection errors. These six parameters form a two-row, three-column affine transformation matrix.

[0026] The parameter error of the affine transformation matrix is ​​corrected through an iterative optimization algorithm, specifically the Levenberg-Marquardt algorithm. The Levenberg-Marquardt algorithm is a numerical optimization algorithm suitable for nonlinear least squares problems. It iteratively adjusts the parameter values ​​of the affine transformation matrix, continuously reducing the reprojection error. The reprojection error is defined as the Euclidean distance between the coordinates of the current frame feature point mapped to the previous frame coordinate system by the currently estimated affine transformation matrix and the corresponding true coordinates of the previous frame feature point. In each iteration, the Levenberg-Marquardt algorithm calculates the Jacobian matrix of the error function with respect to the parameters. The Jacobian matrix contains the partial derivatives of the error with respect to the six affine transformation parameters for each matching point. Based on the magnitude of the error under the current parameters and the Jacobian matrix information, the Levenberg-Marquardt algorithm dynamically adjusts the parameter update step size. When the parameter space is close to the optimal solution, it uses an approximate Gauss-Newton method for fast convergence; when it is far from the optimal solution, it uses gradient descent to ensure stability. The parameter update is obtained by solving a system of linear equations. The coefficient matrix of the linear equations is the transpose of the Jacobian matrix multiplied by the Jacobian matrix, plus a damping factor multiplied by the identity matrix. The damping factor is dynamically adjusted during iteration based on the reduction of error. The iteration process continues until the magnitude of the parameter update is less than a preset tolerance threshold, or the reduction in reprojection error is no longer significant, or the maximum number of iterations is reached. At this point, the affine transformation matrix parameters are considered the optimized result, which can more accurately describe the global motion between frames.

[0027] Motion compensation is performed on subsequent frames using a modified affine transformation matrix. The purpose of motion compensation is to eliminate global jitter and distortion in the image sequence caused by the movement of the UAV platform. The motion compensation process is implemented for each subsequent frame; the frame requiring compensation is called the current frame, and the frame serving as the reference coordinate system is called the reference frame. Motion compensation utilizes an affine transformation matrix optimized by the Levenberg-Marquardt algorithm to reverse-map the coordinates of each pixel in the current frame back to the coordinate system of the reference frame. This reverse mapping is achieved by calculating the inverse transformation of the affine transformation matrix, which transforms the coordinates of the current frame pixel to its corresponding position in the reference frame. This position is typically a sub-pixel precision floating-point coordinate. Image resampling uses bilinear interpolation to obtain the grayscale value of the sub-pixel position. Bilinear interpolation considers the grayscale values ​​of the four nearest neighbor pixels around the transformed coordinate point, and performs a weighted average based on the relative distance between the coordinate point and these four pixels. For an integer coordinate pixel in the current frame, its coordinates in the reference frame coordinate system are calculated using the affine transformation formula, which employs six optimized parameters. Since the transformed coordinates are usually not integers, the corresponding grayscale value in the reference frame is needed to fill the pixel in the current frame. This grayscale value is obtained from the reference frame using bilinear interpolation. Bilinear interpolation first finds the grid containing the transformed coordinate point, which consists of four adjacent integer pixels. The horizontal and vertical distances from the coordinate point to these four pixels are calculated, and the grayscale values ​​of the four pixels are weighted according to distance weights to obtain a weighted estimate of the transformed coordinate point's grayscale value. This grayscale value is assigned to the original integer coordinate pixel in the current frame, completing motion compensation for that point. The above coordinate transformation and bilinear interpolation process is repeated for all pixels in the current frame to generate a compensated image spatially aligned with the reference frame. In the compensated image sequence, the global translation and rotation caused by the drone's flight platform motion are corrected. The static background scene in the sequence remains stable in the image, while independently moving objects in the scene appear to move relative to the stable background. Motion compensation provides a stable spatial reference for subsequent image region segmentation and processing, making the distinction between dynamic and static regions more accurate and reliable.

[0028] Example 3: Region segmentation includes calculating an inter-frame difference map for the motion-compensated image sequence. The inter-frame difference map is obtained by calculating the absolute difference of grayscale values ​​of corresponding pixels in two consecutive motion-compensated images. The calculation process is performed on the t-th and (t-1)-th frames in the sequence. The grayscale values ​​of pixels at the same coordinate positions in the two frames are subtracted, and the absolute value is taken to generate a difference matrix with the same size as the original image. Each pixel value in the difference matrix represents the degree of light intensity change at that position between two adjacent frames; a larger pixel value indicates a more significant change. Regions where pixel changes exceed a threshold are marked. The threshold is set using an adaptive method, dynamically calculated based on the global statistical characteristics of the difference image. A feasible adaptive threshold calculation method is shown in the formula:

[0029] in: This represents the final determined adaptive threshold. This represents the arithmetic mean of all pixel values ​​in the inter-frame difference map. The standard deviation of all pixel values ​​in the inter-frame difference map reflects the dispersion of pixel grayscale changes between adjacent frames. It is an adjustable gain coefficient used to control the sensitivity of the threshold. Adjustable gain coefficient. The value ranges from 0.8 to 2.5, and the specific value is determined based on the complexity of the drone's operating scenario: simple scenarios, such as environmental monitoring in a clear plain under stable lighting conditions with no obvious obstructions. At this point, the threshold selection is more lenient, reducing false positives in static areas; this is suitable for moderately complex scenarios, such as mixed urban and suburban terrain mapping and scenarios with slight changes in lighting. It balances dynamic region detection sensitivity with noise immunity; it is suitable for complex scenarios, such as mountain rescue on cloudy days, high light contrast, and multiple obstructions. At this point, the threshold screening is more stringent, avoiding false labeling of pseudo-dynamic regions caused by noise. By statistically analyzing the dispersion of pixel values ​​in the inter-frame difference image and the standard deviation of all pixel values ​​in the inter-frame difference image, if... This corresponds to a simple scenario. For medium-complex scenarios, For complex scenarios.

[0030] The value of each pixel in the inter-frame difference map is compared with the threshold. Comparison, pixel value greater than Points are marked as foreground points, i.e., areas of change, with pixel values ​​less than or equal to [a certain value]. Points were marked as background points. Dynamic and static regions were distinguished by combining optical flow field analysis results. The optical flow field analysis was calculated using the Lucas-Kanade method. The Lucas-Kanade method assumes that all pixels have a consistent motion vector within a small spatial neighborhood and solves for the motion velocity of each pixel using the least squares method. Optical flow calculation was performed on the motion-compensated image sequence to reduce the interference of global motion on local motion estimation. The change regions marked by the inter-frame difference map were fused with the optical flow field analysis. Within the change regions marked by the inter-frame difference map, those regions with significant and consistent optical flow vectors were identified as true dynamic regions, which typically correspond to moving objects in the scene, such as vehicles or pedestrians. Regions within the change regions marked by the inter-frame difference map with weak, chaotic, or inconsistent optical flow vectors, or those inconsistent with the global motion trend of the background, were identified as pseudo-change regions caused by noise, sudden changes in illumination, or shadow changes. These regions need to be suppressed or classified as static regions in subsequent processing. Morphological processing was performed on the dynamic regions to eliminate mis-segmentation caused by noise. Morphological processing included opening and closing operations. The opening operation performs erosion followed by dilation. Erosion uses a structuring element, such as a 3x3 rectangular kernel, to traverse the binarized dynamic region mask image. The center pixel is retained as the foreground only when all pixels within the structuring element kernel are foreground pixels; otherwise, it is set as a background pixel. Erosion can eliminate small, isolated noise points and smooth the boundaries of dynamic regions. The dilation operation uses the same structuring element. The center pixel is set as the foreground only when at least one pixel within the structuring element kernel is foreground. Dilation can restore the reasonable area of ​​over-eroded dynamic regions. The closing operation performs dilation followed by erosion. Closing fills the holes inside dynamic regions caused by noise or missing textures, making the segmented dynamic region contours more complete and coherent. The morphologically processed binary mask image ultimately and accurately defines the dynamic and static regions in the image.

[0031] Multi-scale noise reduction of dynamic regions includes wavelet decomposition, implemented using discrete wavelet transform (DWT). DWT transforms the image signal of the dynamic region from the spatial domain to the time-frequency domain. DWT uses a set of low-pass and high-pass filters to convolve and downsample the image. The low-pass filter extracts low-frequency approximation information, while the high-pass filter extracts high-frequency detail information. Wavelet decomposition first performs a first-level decomposition on the original dynamic region image, generating four sub-band images: low-frequency approximation sub-band LL1, horizontal high-frequency detail sub-band LH1, vertical high-frequency detail sub-band HL1, and diagonal high-frequency detail sub-band HH1. The low-frequency approximation sub-band LL1 contains the main contours and background information of the image, while the high-frequency detail sub-bands LH1, HL1, and HH1 contain edge, texture, and noise information. To obtain multi-scale analysis capabilities, the low-frequency approximation sub-band LL1 can be recursively decomposed into a second-level decomposition. Using LL1 as input, it is filtered and downsampled again to generate four sub-band images: LL2, LH2, HL2, and HH2. Based on noise reduction requirements and computational resources, deeper decomposition can be performed, with each level analyzing the image signal at a coarser scale. After obtaining sub-images of different frequency bands, threshold filtering is applied to the high-frequency sub-images based on noise statistical characteristics. The noise statistical characteristics assume that the wavelet coefficients in the high-frequency sub-bands are a superposition of signal and noise coefficients, and the noise is typically modeled as additive white Gaussian noise. Threshold filtering uses a soft thresholding function to process each wavelet coefficient in the high-frequency sub-bands. The mathematical expression of the soft thresholding function is: in: Represents the original wavelet coefficients in the high-frequency sub-image. This represents the wavelet coefficients after soft thresholding. It is a threshold. It is a symbolic function. yes The absolute value. Threshold. The choice of threshold is crucial, and a commonly used general threshold estimation method is... ,in This is an estimate of the noise standard deviation, reflecting the noise intensity level. It can be robustly estimated by dividing the absolute value of the median of the wavelet coefficients in the high-frequency subband HH1 by 0.6745. This represents the total number of wavelet coefficients in the high-frequency sub-image. The soft thresholding function selects coefficients whose absolute values ​​are less than the threshold. The wavelet coefficients are set to zero, as these coefficients are considered to be mainly contributed by noise; the absolute values ​​greater than or equal to the threshold are... Wavelet coefficients shrink to zero The coefficients are considered to contain useful signal information but are contaminated by noise. This processing can suppress noise while preserving as much important high-frequency information as possible that represents image edges and details. The reconstructed filtered sub-image retains the details of dynamic regions. The reconstruction process is the inverse process of wavelet decomposition and is called the discrete wavelet inverse transform. The discrete wavelet inverse transform starts from the deepest wavelet decomposition coefficients and performs upsampling and filtering operations on the sub-band images of each decomposition level. Taking the second-level decomposition as an example, the reconstruction process starts from the LL2, LH2, HL2, and HH2 sub-bands. First, the LL2 sub-band is upsampled and zero values ​​are inserted. Then, it is filtered using the reconstruction low-pass filter corresponding to the wavelet decomposition. At the same time, the three high-frequency sub-bands LH2, HL2, and HH2, which have been threshold-filtered, are upsampled and filtered using the reconstruction high-pass filter. The filtered low-frequency components and the three high-frequency components are added together to reconstruct the first-level low-frequency approximate sub-band LL1'. Then, the LL1' subband is subjected to the same upsampling, filtering, and addition operations as the first-level high-frequency subbands LH1, HL1, and HH1, which have undergone threshold filtering, to finally reconstruct the denoised dynamic region image. In the reconstructed image, most of the random noise is effectively suppressed, while the important edges and texture details of moving objects in the dynamic region are preserved during the threshold filtering process, thus achieving a balance between noise reduction and detail preservation.

[0032] Example 4: Texture enhancement processing includes performing Local Binary Pattern Analysis (LDB) on static regions. LDB is an operator used to describe local texture features of an image. LDB defines a fixed-size circular neighborhood centered on each pixel in the static region. The radius of the circular neighborhood determines the scale of the texture analysis. For each sampling point in the circular neighborhood, its gray value is compared with the gray value of the center pixel. If the gray value of the sampling point is greater than or equal to the gray value of the center pixel, the sampling point is marked as 1; otherwise, it is marked as 0. The binary labels of all sampling points in the circular neighborhood are arranged sequentially (e.g., clockwise) to form a binary bit sequence. This binary bit sequence is converted into a decimal number, and the resulting value is the LDB encoding value of the center pixel. The LDB encoding value reflects the texture pattern of the local region centered on that pixel, such as edges, spots, or flat areas. Texture features are extracted by plotting a histogram of the local binary pattern encoded values ​​of all pixels within the entire static region. The horizontal axis of the histogram represents the local binary pattern encoded values ​​(0 to 255 for 8 sampling points), and the vertical axis represents the frequency of each encoded value in the static region. This histogram serves as a feature vector describing the overall texture characteristics of the static region. Information such as the uniformity and complexity of the texture features is derived from the distribution shape of the histogram; for example, a concentrated histogram distribution indicates uniform texture, while a dispersed distribution indicates complex texture.

[0033] The contrast enhancement coefficient is adaptively adjusted based on texture features, determining the intensity of image enhancement. This adaptive adjustment process is based on texture statistics extracted from the local binary pattern histogram, with a key statistic being the texture complexity measure. Texture complexity is quantified by calculating the entropy of the local binary pattern histogram; a higher entropy value indicates a more complex and irregular texture. A mapping relationship is established between the texture complexity measure and the contrast enhancement coefficient, which can be preset as a lookup table. This lookup table defines the appropriate contrast enhancement coefficient value for different texture complexity ranges. Regions with simple, flat textures correspond to lower entropy values; to avoid over-enhancement leading to noise and blockiness, these regions are assigned smaller contrast enhancement coefficients. Regions with complex and rich textures correspond to higher entropy values; these regions can tolerate stronger enhancement without easily producing visual defects, therefore, larger contrast enhancement coefficients are assigned to significantly improve visual clarity. See Table 1 for an example mapping relationship between texture complexity and contrast enhancement coefficient. Table 1: Mapping Relationship between Texture Complexity Ranges and Contrast Enhancement Coefficients Texture complexity interval (based on entropy value calculation) Contrast enhancement coefficient (a) Low complexity [0.0, 2.0) 1.1 Medium-low complexity [2.0, 3.5) 1.3 Medium complexity [3.5, 4.5) 1.5 Medium-high complexity [4.5, 5.5) 1.7 High complexity [5.5, ∞) 1.9 A nonlinear filtering algorithm is employed to enhance the edge structure of static regions. The chosen nonlinear filtering algorithm is adaptive anisotropic diffusion filtering. Adaptive anisotropic diffusion filtering is a filtering technique that adaptively adjusts the smoothing intensity based on local image gradient information. Its core idea is to perform strong smoothing in uniform regions to suppress noise, and weaker or even no smoothing in edge regions to protect the edges. The adaptive anisotropic diffusion filtering process is implemented iteratively, updating the gray value of each pixel in the static region in each iteration. The update amount depends on the gray value difference between the pixel and its neighboring pixels, as well as a function called the diffusion coefficient, which is a function of the local image gradient. The diffusion coefficient function is designed such that its value is close to 1 in regions with small gradients (flat regions), indicating that strong diffusion is allowed for noise reduction; and close to 0 in regions with large gradients (edges), indicating that diffusion is suppressed to protect the edges. A commonly used form of the diffusion coefficient function is the Perona-Malik model, which calculates the diffusion coefficient based on the gradient magnitude. In each iteration, the gray value of a pixel is updated according to the product of its gradient in four directions (east, south, west, and north) and the corresponding diffusion coefficient. The number of iterations needs to be moderate; too few iterations will not result in sufficient noise reduction, while too many iterations may lead to excessive smoothing and loss of detail. After adaptive anisotropic diffusion filtering, noise in static areas is suppressed, while object edges and texture details are preserved or even enhanced, thus improving the visual clarity of the image.

[0034] The specific steps for fusing and reconstructing the processed dynamic and static regions are as follows: Pixel-level alignment is based on the binary mask generated during the region segmentation stage. The dynamic region mask marks the coordinate positions of dynamic pixels, and the static region mask marks the coordinate positions of static pixels. The dynamic and static region images are aligned in the same spatial coordinate system to ensure a one-to-one correspondence between the pixel coordinates of the two regions without spatial offset. Weight allocation adopts an adaptive weighted fusion strategy. The fusion weight for the dynamic region is ω1=0.9-0.95. Since the dynamic region has undergone multi-scale noise reduction processing, it is necessary to prioritize preserving the detail integrity of its moving targets. The fusion weight for the static region is ω2=0.05-0.1. The static region has undergone texture processing... Enhancement is achieved to avoid excessive weighting that leads to distortion due to over-superimposed details, while ensuring ω1+ω2=1. For the boundary pixels between dynamic and static regions, a boundary band with a width of 3-5 pixels is processed using Gaussian weighted averaging. The weight of the boundary pixel is determined by its distance from the center of the region; the closer to the center, the greater the weight; the farther away, the smaller the weight, eliminating jaggedness and brightness abrupt changes at the boundary between the two regions. The pixel value range of the fused image is normalized to ensure that the grayscale value of all pixels is in the range of 0-255, conforming to the standard format of high-definition images, avoiding overexposure (pixel value > 255) or underexposure (pixel value < 0) caused by fusion, ultimately generating an optimized high-definition image sequence.

[0035] Layered encoding of the optimized high-definition image sequence includes performing a discrete cosine transform (DCT) on the sequence, which transforms the image from the spatial domain to the frequency domain. The preset transmission bandwidth conditions refer to a three-level bandwidth threshold standard based on the real-time channel capability of the wireless communication link between the UAV and the ground receiver. Specifically, the high bandwidth condition is a real-time available bandwidth ≥ 10 Mbps, suitable for short-range, unobstructed communication scenarios; the medium bandwidth condition is 5 Mbps ≤ real-time available bandwidth < 10 Mbps, suitable for medium-range, slightly obstructed scenarios; and the low bandwidth condition is real-time available bandwidth < 5 Mbps, suitable for long-range, complex obstructed, or strong electromagnetic interference scenarios. These preset conditions are generated by the system based on the modulation and demodulation method of the UAV communication module, antenna gain parameters, and historical transmission data. Those skilled in the art can directly obtain the real-time bandwidth value through the bandwidth monitoring function of the UAV's onboard communication equipment and match the corresponding conditions. The DCT processes the image in units of image blocks, dividing each frame of the optimized high-definition image sequence into multiple 8-pixel × 8-pixel non-overlapping image blocks. A two-dimensional discrete cosine transform (DCT) is applied to each 8x8 image patch, converting the 64 spatial pixel values ​​in the patch into 64 frequency coefficients. These frequency coefficients include one DC component coefficient and 63 AC component coefficients. The DC component coefficient, located in the upper left corner of the transformed coefficient matrix, represents the average brightness information of the image patch. The AC component coefficients are distributed in other positions of the coefficient matrix, representing detail information at different frequencies within the image patch; coefficients farther from the DC component represent higher-frequency details. The DCT exhibits energy concentration characteristics; for typical natural images, most of the energy after the transformation is concentrated in a few low-frequency coefficients, i.e., the upper left region of the coefficient matrix, while most high-frequency coefficients have amplitudes close to zero. Separating the low-frequency and high-frequency components is achieved by setting a predefined quantization matrix, also an 8x8 matrix, where each element represents the importance weight of the corresponding frequency coefficient. The quantization matrix elements corresponding to the DC component and low-frequency AC components have smaller values, while those corresponding to the high-frequency AC components have larger values. Each coefficient in the discrete cosine transform coefficient matrix is ​​divided by the quantization step size at the corresponding position in the quantization matrix, and then rounded to the nearest integer. Because the quantization step size for low-frequency coefficients is small, the quantization retains high precision and minimal information loss; while the quantization step size for high-frequency coefficients is large, many small-amplitude high-frequency coefficients become zero after division and rounding. In the quantized coefficient matrix, non-zero coefficients are mainly concentrated in the low-frequency region in the upper left corner, while a large number of zero coefficients appear in the high-frequency region in the lower right corner, thus achieving separation of low-frequency and high-frequency information in terms of data volume.

[0036] Low-frequency components are encoded as base layer data, while high-frequency components are encoded as enhancement layer data. Base layer data contains crucial low-frequency information for image reconstruction, extracted from the quantized discrete cosine transform coefficient matrix. Typically, a small region in the upper left corner of the coefficient matrix (e.g., the top 4x4 or 2x2 coefficient submatrix) is defined as base layer data; this region contains the DC component and, most importantly, the low-frequency AC component. Base layer data is compressed using a fine-grained entropy coding algorithm, allowing for the reconstruction of basic image contours and identifiable content even with extremely limited network bandwidth, simply transmitting and decoding base layer data is sufficient. Enhancement layer data contains all non-zero coefficients in the quantized coefficient matrix outside the area covered by base layer data. These coefficients primarily belong to the mid-to-high frequency AC components, providing information on image detail, texture, and edge sharpness. Enhancement layer data can be encoded using relatively coarse-grained entropy coding. When network bandwidth is sufficient, transmitting enhancement layer data along with base layer data can significantly improve the visual quality of the reconstructed image. Entropy coding is a lossless data compression technique based on information theory, used to compress both base and enhancement layer data. Entropy coding is performed independently on the base layer data and enhancement layer data. Entropy coding utilizes the non-uniformity of the probability distribution of symbols in the data, assigning short codewords to symbols with high occurrence probabilities and long codewords to symbols with low occurrence probabilities, thereby reducing the average code length of the overall data. For the discrete cosine transform coefficient matrix, the symbols to be encoded include the amplitude of non-zero coefficients, the run length of zero coefficients, and the block end marker. The encoding process typically consists of two steps. The first step involves a sawtooth scan of the quantized coefficient matrix, converting the two-dimensional coefficient matrix into a one-dimensional symbol sequence. The sawtooth scan starts from the DC component and alternately traverses the coefficient matrix along the diagonal, ensuring that non-zero coefficients appear as early as possible, with zero coefficients concentrated at the end of the sequence. The second step involves variable-length coding of the generated one-dimensional symbol sequence, commonly using Huffman coding. Huffman coding constructs a Huffman code table based on the statistical frequency of each symbol in historical data, representing high-frequency symbols with short binary codes and low-frequency symbols with long binary codes. After entropy coding, the base layer data and enhancement layer data are compressed into a compact binary stream, ready for network transmission.

[0037] See Figure 4First, a discrete cosine transform is performed on the optimized image sequence to separate low-frequency and high-frequency components. Then, entropy coding is used to achieve efficient compression. The upper bar chart shows the difference in the data proportion of the base layer and enhancement layer in hierarchical coding for different image types. The lower bar chart presents the compression ratio for different image types. This chart, through quantitative comparison, intuitively demonstrates the resource allocation and compression efficiency of hierarchical coding under different image types. It provides data support for subsequent strategies to dynamically adjust the transmission priority of the base layer and enhancement layer based on network status. The base layer is the core of image reconstruction and must be prioritized for transmission; the proportion and compression ratio of the enhancement layer determine the cost and benefit of detail transmission, helping to achieve a balance between image quality and transmission efficiency when bandwidth fluctuates.

[0038] Example 5: Dynamically adjusting transmission priorities includes real-time monitoring of network packet loss rate and latency. The network packet loss rate is statistically analyzed using packet sequence numbers and receiver acknowledgment signals. The sender assigns a unique, incrementing sequence number to each transmitted data block. After successfully receiving and verifying the integrity of the data block, the ground receiver returns an acknowledgment signal containing the received sequence number to the sender. The sender maintains a transmission window and a timer, recording the transmission time of each data block. If no acknowledgment signal for the corresponding sequence number is received within a specified time threshold, the data packet is considered lost. The network packet loss rate is calculated as the ratio of the number of lost data packets to the total number of transmitted data packets within a statistical period. Latency is calculated by determining the total time difference between a data packet leaving the sender and receiving an acknowledgment signal from the receiver; this time difference is called the round-trip time. The sender records a precise timestamp when transmitting a data packet and another timestamp when receiving an acknowledgment signal; the difference between the two is the round-trip time of the data packet. The round-trip times of multiple data packets are continuously measured, and their average or minimum value is taken as the current network latency estimate. The minimum value better reflects the inherent propagation delay of the network and is not affected by temporary queuing. Calculating the current channel quality index requires considering both network packet loss rate and latency. The channel quality index is a normalized value used to quantify the instantaneous transmission conditions of the channel. One method for calculating the channel quality index is to map the network packet loss rate and latency to a scoring range of 0 to 1, respectively. The scoring function can be designed as a non-linear function to reflect the non-linear impact of index deterioration. For example, the network packet loss rate scoring function can be set as follows: a score of 1.0 when the packet loss rate is below 1%, linearly decreasing to 0.6 between 1% and 5%, linearly decreasing to 0.3 between 5% and 10%, and a score of 0.1 when the packet loss rate exceeds 10%. The latency scoring function is similar: a score of 1.0 when the latency is below 50 milliseconds, linearly decreasing to 0.5 between 50 and 200 milliseconds, and a score of 0.2 when the latency exceeds 200 milliseconds. The channel quality index is ultimately calculated as the smaller of the network packet loss rate score and the latency score, or their product. Using the minimum value emphasizes the weakest link effect, meaning that a deterioration in either metric will significantly lower the channel quality index. For example, suppose that within a measurement period, the network packet loss rate is 3% and the latency is 80 milliseconds. According to the scoring function, a 3% packet loss rate corresponds to a score of approximately 0.76, and a latency of 80 milliseconds corresponds to a score of approximately 0.825. If the channel quality index is the minimum of the two, then the current channel quality index is 0.76. This index value will be used for subsequent transmission decisions.

[0039] The transmission bandwidth ratio between base layer data and enhancement layer data is dynamically allocated based on the channel quality index (CQI). This bandwidth ratio determines the amount of data each layer is allowed to transmit per unit time. A CQI threshold is set, for example, 0.7. When the calculated CQI is higher than or equal to 0.7, the network is considered to be in good condition. A balanced transmission strategy is then adopted, allocating similar bandwidth ratios to base layer and enhancement layer data, for example, 40% for base layer data and 60% for enhancement layer data, allowing more image detail information to be transmitted for high-quality image reconstruction. When the CQI falls below 0.7, such as 0.76 in the previous example, the system determines that network congestion is beginning to appear. At this point, a congestion control mechanism is triggered, dynamically adjusting bandwidth allocation and significantly increasing the transmission priority of base layer data. The specific bandwidth allocation can be linearly or piecewise linearly interpolated based on the CQI value. For example, when the CQI is 0.7, the bandwidth allocation for base layer data is 70%, and for enhancement layer data, it is 30%; when the CQI drops to 0.4, the bandwidth allocation for base layer data increases to 95%, and for enhancement layer data, it is only 5%. For a channel quality index (CQI) of 0.76, interpolation calculations may adjust the bandwidth allocation for base layer data to approximately 65%, while reducing the allocation for enhancement layer data to approximately 35%. In extreme network congestion scenarios, such as when the CQI falls below 0.3, the transmission strategy may completely halt the transmission of enhancement layer data, dedicating all available bandwidth to ensuring reliable transmission of base layer data. This guarantees that the ground receiver can receive at least one identifiable but less detailed version of the base image. This dynamic adjustment mechanism ensures robustness and adaptability of image transmission under varying network conditions.

[0040] Sending the adjusted base layer and enhancement layer data to the ground receiver via a multipath transmission protocol involves splitting the base layer and enhancement layer data into multiple data blocks. The size of each data block is set according to the maximum transmission unit (MTU) of the network path to avoid fragmentation. For example, the base layer and enhancement layer streams are each divided into 1400-byte data blocks, with each block appended a header containing a sequence number, path identifier, and data type. Each data block is assigned an independent transmission path identifier, specifying which physical network link the data block should be transmitted through. The UAV platform may be equipped with multiple network interfaces, such as a 4G / 5G cellular network interface and a satellite communication interface. The transmission path identifiers are bound to these network interfaces; path identifier 1 corresponds to the 4G / 5G link, path identifier 2 corresponds to the satellite link, and so on. The multipath transmission protocol maintains a path status table, recording the current available bandwidth, packet loss rate, and latency for each path. Based on the dynamically adjusted bandwidth allocation ratio, the multipath transmission protocol schedules the base layer and enhancement layer data blocks onto different paths. When network congestion occurs, basic layer data blocks are prioritized for routing to paths with more stable quality and lower latency, such as 4G / 5G links; while enhancement layer data blocks may be routed to satellite links with potentially higher bandwidth but also higher latency, or in some cases, their transmission may be temporarily suspended.

[0041] A redundancy check mechanism ensures the complete reception of data blocks. Before data transmission, the redundancy check mechanism calculates an error detection code for each data block. The error detection code is calculated using the Cyclic Redundancy Check (CRBC) algorithm. The CRBC algorithm treats the data block to be transmitted as a long binary number, divides this binary number by a predefined polynomial, and uses the remainder as the CRBC code. The CRBC code is appended to the end of the data block and transmitted along with it. Upon receiving the data block, the ground receiver performs the same division operation on the received data using the same generator polynomial as the sender. If the remainder is zero, the data block is considered error-free during transmission, and the receiver returns an acknowledgment signal for the data block's sequence number to the sender. If the remainder is not zero, it indicates a bit error in the data block. The receiver discards the erroneous data block and returns a negative acknowledgment signal to the sender, requesting retransmission of the data block. The processing logic of a multipath transport protocol can identify retransmission requests for a specific data block and, based on real-time path status information, select the path that is currently considered the most reliable to retransmit the data block, rather than necessarily using the original transmission path.

[0042] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for processing and transmitting high-definition images based on unmanned aerial vehicles, characterized in that, The method comprises the following steps: Collecting a plurality of high-definition image sequences obtained during the flight of the unmanned aerial vehicle, the high-definition image sequences containing image data at different time stamps; Performing frame-by-frame analysis on the high-definition image sequences, extracting key feature points in each frame of image, and generating image feature descriptors according to the key feature points; Establishing a feature matching relationship between adjacent frames based on the image feature descriptors, and calculating inter-frame motion vectors; Constructing a global motion model according to the inter-frame motion vectors, and performing motion compensation on the high-definition image sequences through the global motion model; Segmenting the motion-compensated high-definition image sequences into dynamic regions and static regions; Performing multi-scale noise reduction processing on the dynamic regions, and simultaneously performing texture enhancement processing on the static regions; Fusing and reconstructing the processed dynamic regions and static regions to generate an optimized high-definition image sequence; According to a preset transmission bandwidth condition, the optimized high-definition image sequence is encoded in layers to generate base layer data and enhancement layer data; Based on the network state, the transmission priorities of the base layer data and the enhancement layer data are dynamically adjusted; The adjusted base layer data and enhancement layer data are sent to the ground receiving end through a multi-path transmission protocol. 2.The method of claim 1, wherein, The extraction of key feature points in each frame of image comprises: An edge detection algorithm is used to identify contour information in the image, and feature points with significant gradients are selected according to the contour information; Local descriptor calculation is performed on the feature points to generate feature vectors with rotation invariance; According to the similarity of the feature vectors, the feature points are clustered, and redundant feature points are removed. 3.The method of claim 2, wherein, The establishment of a feature matching relationship between adjacent frames based on the image feature descriptors comprises: The Euclidean distance between the feature descriptors of the current frame and the previous frame is calculated to determine the initial matching pairs; The initial matching pairs are filtered through a random sample consensus algorithm to remove mis-matching points; The inter-frame motion vectors are calculated based on the filtered matching points.

4. The method of claim 3, wherein the method further comprises: The construction of a global motion model according to the inter-frame motion vectors comprises: An affine transformation matrix is fitted using the inter-frame motion vectors to describe the global motion trend between frames; The parameter error of the affine transformation matrix is corrected through an iterative optimization algorithm; The subsequent frames are motion-compensated using the corrected affine transformation matrix.

5. The method of claim 4, wherein, The segmentation of the motion-compensated high-definition image sequences comprises: The inter-frame difference map of the motion-compensated image sequences is calculated to mark the regions where the pixel changes exceed the threshold value; The dynamic regions and static regions are distinguished based on the results of the optical flow field analysis; Morphological processing is performed on the dynamic regions to eliminate false segmentation caused by noise. 6.The method of claim 5, wherein, The multi-scale noise reduction processing of the dynamic regions comprises: Wavelet decomposition is performed on the dynamic regions to obtain sub-images of different frequency bands; Threshold filtering is performed on the high-frequency sub-images according to the noise statistical characteristics; The filtered sub-images are reconstructed to retain the detail information of the dynamic regions.

7. The method of claim 6, wherein the method further comprises: The simultaneous texture enhancement processing of the static regions comprises: The static region is subjected to local binary pattern analysis, texture features are extracted, a contrast enhancement coefficient is adaptively adjusted according to the texture features, and a nonlinear filtering algorithm is used to strengthen the edge structure of the static region. 8.The method of claim 7, wherein, The layered encoding of the optimized high-definition image sequence comprises: Discrete cosine transform is performed on the optimized high-definition image sequence to separate low-frequency and high-frequency components; The low-frequency components are encoded as the base layer data, and the high-frequency components are encoded as the enhancement layer data; Entropy encoding algorithm is used to compress the base layer data and the enhancement layer data. 9.The method of claim 8, wherein, The dynamic adjustment of the transmission priority of the base layer data and the enhancement layer data based on the network state comprises: The network packet loss rate and delay time are monitored in real time, and the current channel quality index is calculated; The transmission bandwidth proportion of the base layer data and the enhancement layer data is dynamically allocated according to the channel quality index, and the base layer data is preferentially transmitted when the network is congested. 10.The method of claim 9, wherein, The adjusted base layer data and enhancement layer data are sent to the ground receiving end through the multi-path transmission protocol, which comprises: The base layer data and the enhancement layer data are split into multiple data blocks, each data block is assigned an independent transmission path identifier, and the integrity of the data block is ensured through a redundancy check mechanism.

Images