Unmanned aerial vehicle aerial photography area image quality optimization method based on image processing

Abstract

The invention relates to the technical field of unmanned aerial vehicle aerial image intelligent processing, and discloses an unmanned aerial vehicle aerial area image quality optimization method based on image processing. According to the method, an original image stream and a flight state data stream are synchronously received, various degradation factors are identified and quantized through perceptual degradation analysis, and feature descriptors containing degradation components and weights are generated. Based on the descriptor, an adaptive search topology network which takes image quality expectation as traction and can reflect a quality evolution path is dynamically constructed. In the network, multi-stage guided focused retrieval is executed according to the degradation type and intensity, and the most relevant target image unit is locked to form a to-be-processed cluster. And implementing a parameter-adaptive hybrid enhancement strategy on the cluster, correcting and compensating for different degradation components, generating an optimized image, and feeding back to update the network. According to the method, precise diagnosis and self-adaptive optimization processing of aerial image hybrid degradation are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent processing technology for drone aerial images, specifically a method for optimizing the quality of drone aerial images of a specific area based on image processing. Background Technology

[0002] Existing methods for optimizing UAV aerial image quality primarily rely on pre-defined general image enhancement algorithms or independent analysis of single-frame images. These techniques typically perform operations such as denoising, sharpening, or color enhancement based on pixel-domain statistical features or frequency domain transformations. However, in dynamic flight operations, image degradation is caused by a combination of factors, including atmospheric scattering, sensor noise, platform vibration, and motion blur. Existing methods lack the ability to collaboratively analyze multi-source heterogeneous degradation factors, failing to effectively separate and quantify the specific contributions of each degradation component. This leads to fixed or blindly adjusted optimization algorithm parameters, making it difficult to adapt to complex and changing imaging environments and resulting in unstable processing effects.

[0003] Existing technologies often select reference data from image databases based on image content feature matching or global quality score ranking. These retrieval mechanisms have weak correlations with the specific type and intensity of image degradation, failing to establish an intrinsic quality correlation model between image database units. The retrieval results are mostly single images with similar visual content or high quality scores, unable to provide targeted processing clues or state references for the current specific, mixed degradation pattern, thus lacking precise guidance for subsequent optimization processes.

[0004] A technical solution is needed that can accurately quantify the causes of hybrid degradation in aerial images and, based on the diagnostic results, dynamically retrieve the processing criteria that best match the current degradation state from historical data, thereby driving an adaptive and precise image quality optimization process. Summary of the Invention

[0005] The purpose of this invention is to provide a method for optimizing the image quality of drone aerial photography areas based on image processing, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for optimizing the image quality of an area captured by a drone based on image processing, the method comprising: Receive raw image streams and flight status data streams synchronously collected by the UAV platform during flight operations; Perceptual degradation analysis is performed on the original image stream to identify multiple degradation factors, including atmospheric disturbances, inherent noise of the imaging device, and geometric distortion caused by flight conditions. The degree of degradation caused by each degradation factor to the current image stream is quantified, and a degradation feature descriptor containing multiple degradation components and their weights is constructed. Based on the degradation feature descriptor, an adaptive search structure covering the target region and driven by image quality expectations is dynamically constructed. The adaptive search structure organizes the image units in the image library into a topological network that can reflect the quality evolution path. In the adaptive search structure, a multi-level guided focused retrieval is performed based on the degradation type and intensity indicated by the degradation feature descriptor, filtering and locking a series of target image units most relevant to the degradation feature descriptor layer by layer to form a cluster of units to be processed; The cluster of units to be processed is subjected to quality optimization processing. The quality optimization processing adopts a hybrid enhancement strategy based on the parameter adaptation of the degradation feature descriptor. Correction and compensation operations are applied to different degradation components to generate optimized image units. The optimized image units are then fed back into the adaptive search structure to update their network state.

[0007] Preferably, the perceptual degradation analysis of the original image stream identifies multiple degradation factors, including atmospheric disturbances, inherent noise of the imaging device, and geometric distortion caused by flight conditions, and quantifies the degree of degradation caused by each degradation factor to the current image stream, thereby constructing a degradation feature descriptor containing multiple degradation components and their weights, including: A continuous frame sequence is extracted from the original image stream. By analyzing the inter-frame differences and intra-frame statistical characteristics, dynamic degradation components and static degradation components are initially separated. The attitude angular velocity and position jitter information in the flight status data stream are spatiotemporally aligned and correlated with the dynamic degradation components to calculate the specific parameters of the three degradation factors caused by flight motion: motion blur, geometric distortion, and field of view instability. Multi-scale texture analysis and noise power spectrum estimation are performed on single-frame images in the original image stream to separate the influence models of three degradation factors: fixed pattern noise determined by the characteristics of the imaging sensor, random noise, and optical attenuation caused by atmospheric scattering. The preliminary results after correlation of flight status data stream and single-frame analysis are cross-validated to quantify the confidence and influence weight of each degradation factor. The parameters, influence models and their confidence and weight of all degradation factors are integrated into a structured data object, which is the degradation feature descriptor.

[0008] Preferably, the step of dynamically constructing an adaptive search structure covering the target region based on degraded feature descriptors and driven by image quality expectations includes: The metadata and shallow visual features of all image units in the aerial image library are read in advance. Based on the acquisition time, geographical location and basic quality score of the image units, a preliminary spatial-quality correlation mapping table is established. Guided by the dominant degradation type in the degradation feature descriptor, image units that have been historically affected by similar degradation problems and have shown significant optimization effects are selected from the spatial-quality correlation mapping table and used as the initial seed nodes for constructing the adaptive search structure. Using the initial seed node as the core, directed connections are established between nodes based on the content coherence, quality inheritance relationship and spatiotemporal proximity between image units. The direction of the connection represents the direction that quality optimization may be passed or referenced, and the strength of the connection is determined by the similarity of features between nodes and the correlation of quality gradient. An online update mechanism is introduced, which adjusts the connection strength and topological relationship of nodes in the adaptive search structure in real time based on the latest received original image stream and its corresponding degradation feature descriptor. For newly emerging degradation patterns, the online update mechanism creates new node branches in the adaptive search structure to accommodate and represent the degradation patterns.

[0009] Preferably, in the adaptive search structure, a multi-level guided focused retrieval is performed based on the degradation type and intensity indicated by the degradation feature descriptor, progressively filtering and locking a series of target image units most relevant to the degradation feature descriptor, forming a cluster of units to be processed, including: The degenerate feature descriptor is encoded into a query vector that the adaptive search structure can understand, and the query vector is input into the entry node of the adaptive search structure; The first-level search is initiated. Within the direct neighborhood of the entry node, the matching degree between the query vector and the feature vector of the image unit represented by each neighboring node is calculated. Nodes with matching degrees exceeding the primary threshold are selected to form the primary candidate set. The second-level search is initiated, with each node in the primary candidate set as a new search starting point, and the search is explored downstream of the adaptive search structure. The exploration depth is limited by a preset number of hops. During the downstream exploration, not only is the direct matching degree between the query vector and the downstream node calculated, but the cumulative effect of the connection strength on the path from the primary candidate node to the downstream node is also evaluated. The scores of both are combined to select secondary candidate nodes. All secondary candidate nodes are deduplicated and merged, sorted according to their comprehensive scores, and the set of nodes that rank high and meet the final threshold is selected. The image unit corresponding to the set of nodes is locked as the target image unit. All target image units together constitute the unit cluster to be processed. At the same time, the complete retrieval path from the entry node to each target image unit is recorded.

[0010] Preferably, the quality optimization process performed on the cluster of units to be processed employs a hybrid enhancement strategy based on parameter adaptation according to the degenerate feature descriptor, including: The attributes of each target image unit in the cluster to be processed are analyzed, and the retrieval path information associated with the target image unit is extracted from the adaptive search structure; Based on the parameters for motion blur degradation components in the degradation feature descriptor, deconvolution filtering with specific direction and intensity is applied to the target image unit; based on the parameters for geometric distortion degradation components in the degradation feature descriptor, grid-based elastic deformation correction is performed on the target image unit. Based on the influence model of optical attenuation degradation component in the degradation feature descriptor, atmospheric light curtain estimation and contrast restoration operations are performed on the target image unit; Based on the models and weights of various noise degradation components in the degradation feature descriptor, the target image unit is processed by a combination of adaptive filtering and wavelet domain threshold denoising. The intermediate results after the correction and compensation operations for motion blur, geometric distortion, optical attenuation and various types of noise are fused together. During the fusion process, the fusion coefficients are assigned according to the weights of different degradation components to generate the final optimized image unit.

[0011] Preferably, feeding the optimized image units back into the adaptive search structure to update their network state includes: Calculate the quality difference measure between the optimized image unit and the original target image unit in the unit cluster to be processed, and the absolute quality score of the optimized image unit; Based on the quality difference metric and absolute quality score, a new node is created for the optimized image unit in the adaptive search structure, or the feature representation of the node corresponding to the original target image unit is updated. Based on the effectiveness of the optimization process and the retrieval path, the strength of all connections on the path from the entry node to the new or updated node in the adaptive search structure is adjusted. Connections that lead to successful optimization are strengthened, while invalid or inefficient connections are weakened. The newly generated optimized image units and their metadata are archived into the aerial image library, and the spatial-quality association mapping table is updated synchronously.

[0012] Preferably, the step of processing the target image unit using a combination of adaptive filtering and wavelet domain thresholding based on the model and weights of various noise degradation components in the degradation feature descriptor includes: Extract the noise model parameters and their respective influence weights for fixed-pattern noise and random noise from the degradation feature descriptor; Based on the model parameters of the fixed-pattern noise, the spatial distribution pattern of the noise is located on the target image unit, and a matching spatial variation filter is constructed. The kernel function of the spatial variation filter is adaptively adjusted according to the local statistical characteristics of the noise pattern to suppress the fixed-pattern noise. Based on the model parameters of the random noise, its noise power spectrum characteristics are determined, and the optimal number of decomposition layers for wavelet transform is calculated, as well as different threshold functions are set for different sub-bands. Wavelet transform is performed on the image after spatial variation filtering to obtain high-frequency subband coefficients at multiple scales and directions; Based on the threshold function, each high-frequency subband coefficient is subjected to soft or hard thresholding. Coefficients with amplitudes below the threshold are treated as noise and suppressed, while coefficients with amplitudes above the threshold are retained to maintain image details. The high-frequency subband coefficients after thresholding are subjected to inverse wavelet transform to reconstruct the image after removing random noise. Based on the influence weights of fixed-pattern noise and random noise extracted from the degraded feature descriptor, the output of spatial variation filtering and the output of wavelet threshold denoising are weighted and fused to generate the final noise-suppressed image.

[0013] Preferably, the perceptual degradation analysis further includes an environmental context modeling step, comprising: Collect flight environment data synchronized with the original image stream, the flight environment data including at least light intensity, atmospheric visibility and temperature and humidity; An empirical correlation model between environmental data and image degradation factors is established, and the degradation factor parameters calculated by image content analysis and flight status correlation are calibrated using the empirical correlation model for environmental context calibration. The calibrated parameters are used as more accurate inputs to update and correct degenerate feature descriptors.

[0014] Preferably, the dynamic construction process of the adaptive search structure further includes a structure self-optimization step, including: Regularly evaluate the retrieval efficiency and accuracy of the adaptive search structure in historical queries; When the retrieval efficiency or accuracy is lower than the set standard, the structure self-optimization step is triggered. The structure self-optimization step iteratively adjusts the node division rules, connection establishment criteria, and connection strength calculation formula of the adaptive search structure based on historical query logs and optimization result feedback. While ensuring compatibility with existing data, the adjusted new rules are applied to the adaptive search structure to complete the reconstruction of its topology and the reconfiguration of its parameters.

[0015] Preferably, the multi-level guided focused retrieval process further includes a retrieval path backtracking verification step, including: After locking onto the target image unit and forming a cluster of units to be processed, the system automatically performs reverse simulation verification along the recorded complete retrieval path; Simulation verification is performed by calculating whether the same target image unit can still be stably retrieved along the same path, assuming that the current degraded feature descriptor is known. If the success rate of the simulation verification is lower than the preset value, the search path will be marked as an unstable path, and the dependence weight on the search path or its similar paths will be reduced in subsequent searches.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By simultaneously analyzing the raw image stream and flight status data stream, mixed degradation factors such as atmospheric disturbance, equipment noise, and motion geometry distortion are identified and quantified, constructing a degradation feature descriptor containing multiple components and their weights. This enables a refined and structured diagnosis of the causes of image degradation, transforming vague sensory quality assessments into resolvable physical or mathematical model parameters. Based on this descriptor, the degree of influence of different degradation sources can be clearly defined, providing precise operational targets for subsequent processing. This allows enhancement algorithms to move beyond general global adjustments to the entire image, enabling targeted correction and compensation operations with precisely tunable parameters for different degradation components.

[0017] A dynamic topological network, driven by image quality expectations and reflecting the path of quality evolution, is constructed to organize units in the image database into an interconnected state graph. Based on degradation feature descriptors, a multi-level guided focused retrieval is performed within this network. The retrieval logic prioritizes images with similar visual content over those with the most relevant quality state nodes, based on degradation type and intensity. This method can identify a cluster of units from massive datasets that best match the specific degradation problem, representing effective paths or sample sets for solving such quality issues. The optimized results are fed back into the network to update its state, allowing the search structure to continuously accumulate knowledge and evolve its ability to model the relationship between complex degradation problems and optimal solutions, forming an optimization loop with learning and evolutionary characteristics. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the image quality optimization method for UAV aerial photography areas based on image processing described in this invention. Figure 2 A flowchart for perceptual degradation parsing and descriptor construction; Figure 3 A flowchart for dynamically constructing an adaptive search structure; Figure 4 A bar chart comparing the scores before and after optimizing the quality of drone aerial images; Figure 5 A trend chart of quality scores for phased optimization of drone aerial images. 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 optimizing the image quality of an area captured by a UAV (Unmanned Aerial Vehicle) based on image processing. The method includes: receiving a raw image stream and a flight status data stream synchronously acquired by a UAV platform during flight operations; performing perceptual degradation analysis on the raw image stream to identify multiple degradation factors, including atmospheric disturbances, inherent noise of the imaging equipment, and geometric distortion caused by flight status, and quantifying the degree of degradation caused by each degradation factor to the current image stream, thereby constructing a degradation feature descriptor containing multiple degradation components and their weights; dynamically constructing an adaptive search structure covering the target area and driven by image quality expectations based on the degradation feature descriptor, which organizes image units in the image library into a topological network that reflects the quality evolution path; and performing a multi-level guided focused retrieval based on the degradation type and intensity indicated by the degradation feature descriptor, progressively filtering and locking a series of target image units most relevant to the degradation feature descriptor, forming a cluster of units to be processed. The quality optimization process is performed on the cluster of units to be processed. This quality optimization process adopts a hybrid enhancement strategy based on the parameter adaptation of the degradation feature descriptor. Correction and compensation operations are applied to different degradation components to generate optimized image units. The optimized image units are then fed back into the adaptive search structure to update their network state.

[0021] In one embodiment of the present invention, see [reference] Figure 2In practice, perceptual degradation analysis of the original image stream involves a series of sequentially cooperating steps. First, a continuous frame sequence is extracted from the original image stream. By analyzing the inter-frame differences and intra-frame statistical characteristics of the continuous frame sequence, dynamic degradation components that change over time and relatively stable static degradation components are initially separated from the image. In a more detailed implementation, attitude angular velocity and position jitter information from the flight status data stream are extracted and aligned with the aforementioned dynamic degradation components using high-precision timestamps and spatial position mapping. Through correlation mapping, specific parameters of motion blur degradation factors, geometric distortion degradation factors, and field-of-view instability degradation factors directly caused by flight motion are calculated, such as the point spread function and blur direction angle of motion blur. In some embodiments, multi-scale texture analysis and noise power spectrum estimation are simultaneously performed on single-frame images in the original image stream. Multi-scale texture analysis distinguishes image content from degradation patterns by examining local statistical regularities of the image at different resolutions, while noise power spectrum estimation quantifies the energy distribution of noise in the frequency domain. These two analyses work together to separate mathematical models of fixed-pattern noise degradation factors, random noise degradation factors, and optical attenuation degradation factors caused by atmospheric scattering, all determined by the characteristics of the imaging sensor. In specific implementations, the preliminary results after correlation of flight status data streams and single-frame analysis are cross-validated. Cross-validation assesses consistency by comparing the estimation results of different methods for the same degradation phenomenon, thereby quantifying the confidence level of each degradation factor and its weight in influencing overall image quality. Finally, the parameters, influence models, confidence levels, and weights of all degradation factors are integrated into a structured data object, which is defined as a degradation feature descriptor.

[0022] In practical implementation, the quality optimization process employs a hybrid enhancement strategy based on parameter adaptation using degradation feature descriptors. The processing of noise degradation components combines adaptive filtering with wavelet domain thresholding. Specifically, the specific noise model parameters for fixed-pattern noise degradation factors and random noise degradation factors, along with their pre-quantized influence weights, are extracted from the constructed degradation feature descriptors. Optionally, based on the model parameters of the fixed-pattern noise degradation factor, the spatial distribution pattern of the fixed-pattern noise is located on the target image unit. This spatial distribution pattern refers to the periodic or regular pattern exhibited by the noise on the image sensor pixels. A matching spatial variation filter is then constructed, and its kernel function is adaptively adjusted based on the local statistical characteristics of the fixed-pattern noise pattern to suppress it. In practical implementation, the noise power spectrum characteristics of the random noise are determined based on the model parameters of the random noise degradation factor. The optimal decomposition level for subsequent wavelet transform is calculated based on these characteristics, and different threshold functions are set for different frequency sub-bands. It can be understood that performing a wavelet transform on the image after spatial variation filtering yields high-frequency sub-band coefficients at multiple scales and directions. Based on a preset threshold function, each high-frequency subband coefficient is subjected to soft or hard thresholding. Coefficients with amplitudes below the threshold are treated as noise and suppressed, while coefficients with amplitudes above the threshold are retained to maintain image edges and texture details. In a specific implementation, the high-frequency subband coefficients after thresholding are subjected to inverse wavelet transform to reconstruct the image after removing random noise. Optionally, the output of spatial variation filtering and the output of wavelet thresholding are weighted and fused according to the influence weights of fixed-pattern noise degradation factors and random noise degradation factors extracted from the degradation feature descriptors to generate the final noise-suppressed image. In some embodiments, the threshold function used for wavelet domain thresholding is designed as a nonlinear form related to the subband coefficient distribution, and its formula is expressed as: in: Indicates the first The decomposition scale, the first The threshold applied to each direction subband. This represents the estimate of the random noise standard deviation obtained from the degraded feature descriptor. This indicates the total number of pixels in the currently processed image block. It is an adjustment parameter related to the image content. This threshold design allows for a stricter threshold to be applied in the high-frequency subband to remove noise, while the threshold is appropriately relaxed in the lower-frequency subband to retain more detail.

[0023] In one embodiment of the present invention, see [reference] Figure 3In specific implementation, the metadata and shallow visual features of all image units in the aerial image database are pre-read. The metadata includes the acquisition timestamp, GPS coordinates, UAV platform model, and sensor parameters. The shallow visual features include the image's average brightness, contrast, edge density, and color histogram statistics. Based on the image unit's acquisition time, geographical location, and basic quality score calculated by a preset algorithm, a preliminary spatial-quality association mapping table is established. This spatial-quality association mapping table is a query structure that links spatial grid indexes, time windows, and image unit quality evaluation. In some embodiments, the dominant degradation type in the degradation feature descriptor is used as the query guide. The dominant degradation type is determined based on the degradation component with the highest weight in the degradation feature descriptor, such as motion blur degradation factor or optical attenuation degradation factor. Image units that have historically been affected by similar degradation problems and have shown significant optimization effects are selected from the spatial-quality association mapping table. The selection criteria include that the quality improvement of the image unit for the same degradation factor in its historical optimization records exceeds a preset threshold. These image units that meet the criteria are used as the initial seed nodes for constructing an adaptive search structure. In the implementation, the initial seed node is used as the core of network construction. Directed connections are established between nodes based on the content coherence, quality inheritance relationship, and spatiotemporal proximity among image units. Content coherence is measured by calculating the matching degree of scale-invariant feature transform descriptors between image units. Quality inheritance relationship refers to the derivational relationship between new image units generated after optimization of an image unit. Spatiotemporal proximity is calculated based on the acquisition time difference and geographical distance of image units. The direction of the directed connection represents the direction in which quality optimization may be passed on or referenced, such as from a low-quality image unit to a high-quality image unit generated by its optimization. The strength of the directed connection is determined by the similarity of features between nodes and the correlation of quality gradients. In the implementation, an online update mechanism is introduced. This mechanism adjusts the connection strength and topological relationship of nodes in the adaptive search structure in real time based on the latest received original image stream and its corresponding degradation feature descriptor. For newly emerging degradation patterns, the online update mechanism creates new node branches in the adaptive search structure to accommodate and represent the new degradation patterns. It is understandable that the perceptual degradation analysis also includes an environmental context modeling step, which is executed in parallel or sequentially with the aforementioned image content analysis and flight state association calculation.

[0024] In specific implementations, flight environment data is collected synchronously with the original image stream. This flight environment data includes at least light intensity, atmospheric visibility, and temperature and humidity data collected by environmental sensors mounted on the UAV platform. This data has a strict time synchronization relationship with each frame of the original image. In some embodiments, an empirical correlation model is established between environmental data and image degradation factors. This empirical correlation model is trained using historical datasets and describes, for example, the functional relationship between atmospheric visibility and the optical attenuation coefficient, as well as the impact of changes in light intensity on image noise levels. In specific implementations, the empirical correlation model is used to perform environmental context calibration on the degradation factor parameters calculated by image content analysis and flight state correlation. Environmental context calibration refers to using real-time environmental data as input, calculating a calibration factor through the empirical correlation model, and using this calibration factor to correct the initially calculated degradation parameters. Optionally, the calibrated parameters are used as more accurate input to update and correct the parameters and weights of the corresponding degradation components in the previously constructed degradation feature descriptor. It can be understood that the output of the environmental context modeling step is a more accurate degradation feature descriptor after being corrected by the environmental data. This corrected degradation feature descriptor will be used to subsequently build an adaptive search structure or perform quality optimization processing. In practice, the environmental context calibration process can be implemented through a parameterized function, the formula of which is expressed as: in: Indicates the first after environmental context calibration The parameters of each degradation factor. This indicates the preliminary solution obtained through image content analysis and flight status correlation. The original parameters of each degradation factor. This represents an empirical coefficient related to the type of degradation factor, used to control the intensity of environmental impacts. Represents an environment-dependent data vector and historical degradation pattern statistics The mapping function. It is an environmental data vector that includes dimensions such as light intensity, atmospheric visibility, temperature and humidity. This indicates the relationship between historical data and current environmental conditions. In similar cases, the first The statistical distribution characteristics of each degradation factor. This formula describes how to use environmental data and historical knowledge to scale and adjust the parameters of the initially calculated degradation factors to achieve environmental context calibration.

[0025] In one embodiment of the present invention, in a specific implementation, the degenerate feature descriptor is encoded into a query vector that the adaptive search structure can understand and process. The query vector is a numerical representation whose dimension is consistent with the dimension of the node feature vector in the adaptive search structure. The query vector is input to a preset entry node of the adaptive search structure. The entry node is a starting node selected from the adaptive search structure network based on the spatiotemporal attributes or task type of the current processing task. It can be understood that, upon initiating the first-level retrieval, within the direct neighborhood of the entry node (where the direct neighborhood refers to the set of all nodes directly connected to the entry node through directed connections), the matching degree between the query vector and the feature vector of the image unit represented by each neighboring node is calculated. The matching degree is achieved by calculating the cosine similarity or the reciprocal of the Euclidean distance. Nodes with matching degrees exceeding a preset primary threshold are selected to form a primary candidate set. In some embodiments, a second-level search is initiated, using each node in the primary candidate set as a new starting point to explore downstream of the adaptive search structure. The downstream direction is determined based on the direction of directed connections, and the exploration depth is limited by a preset number of hops. During the downstream exploration, not only is the direct matching degree between the query vector and the downstream node calculated, but the cumulative effect of the connection strength along the path from the primary candidate node to the downstream node is also evaluated. By combining the direct matching degree score and the cumulative path strength score, secondary candidate nodes are selected. In a specific implementation, all secondary candidate nodes are deduplicated and merged, and sorted in descending order according to their comprehensive scores. The set of nodes with the highest ranking and meeting the preset final threshold is selected. The image units corresponding to the node set are locked as target image units. All target image units together constitute the unit cluster to be processed. At the same time, the complete search path from the entry node to each target image unit is recorded. The complete search path is a sequence of nodes and directed connections. Optionally, the dynamic construction process of the adaptive search structure also includes a structure self-optimization step.

[0026] In practice, the structure self-optimization stage periodically evaluates the retrieval efficiency and accuracy of the adaptive search structure in historical queries. Retrieval efficiency is measured by average query response time or computational resource consumption, while accuracy is evaluated by the effectiveness of quality improvement generated by the target image units locked by historical queries in subsequent quality optimization processing. In some embodiments, when the retrieval efficiency or accuracy falls below a set standard, the structure self-optimization stage is triggered. Based on historical query logs and optimization result feedback, the structure self-optimization stage iteratively adjusts the node partitioning rules, connection establishment criteria, and connection strength calculation formula of the adaptive search structure. The node partitioning rules determine how to map or create a new image unit as a node in the adaptive search structure. The connection establishment criteria define the conditions under which a directed connection is established between two nodes. The connection strength calculation formula is used to quantify the weight values ​​of directed connections. In practice, while ensuring compatibility with existing data, the iteratively adjusted new rules are applied to the adaptive search structure to complete the reconstruction of the adaptive search structure's topology and the reconfiguration of its parameters. Reconstruction may involve merging, splitting, or re-clustering nodes, while reconfiguration involves updating the strength values ​​of all directed connections. It is understandable that adjusting the connection strength calculation formula is one of the core aspects of the structural self-optimization process. Its purpose is to enable the adaptive search structure to more accurately reflect the potential relationships of quality evolution between image units. An adjustable connection strength calculation formula is expressed as follows: in: Indicates from node Pointing to node The directed connection strength. Represents a node With nodes Feature vector of the corresponding image unit and The similarity between them. Represents a node With nodes The quality gradient vector of the corresponding image unit in the historical optimization process and The correlation between them. It is a balancing factor between 0 and 1, used to adjust the relative importance of feature similarity and quality gradient correlation in connection strength calculation. It is a composite distance function used to calculate nodes. (Collection time is) The geographical location is ) and nodes (Collection time is) The geographical location is The spatiotemporal distance between them. It is a damping coefficient used to control the attenuation effect of spatiotemporal distance on connection strength. The structural self-optimization process, through analysis of historical data, iteratively optimizes the balance factor in the formula. and damping coefficient The value of and the adjustment of the similarity function and correlation function The specific calculation method is used to improve the overall performance of the adaptive search structure.

[0027] In one embodiment of the present invention, in a specific implementation, complete retrieval path information associated with the target image unit is extracted from the adaptive search structure. This complete retrieval path information records the sequence of nodes and directed connections traversed from the entry node to the locked target image unit. This information is used to understand the contextual logic of the target image unit being selected. It can be understood that, based on the parameters for the motion blur degradation component in the degradation feature descriptor, a deconvolution filter with specific direction and intensity is applied to the target image unit. The parameters of the motion blur degradation component include the size, shape, and angle of the point spread function. The deconvolution filter constructs an iterative or regularized deconvolution kernel based on these parameters for image restoration. In some embodiments, based on the parameters for the geometric distortion degradation component in the degradation feature descriptor, a grid-based elastic deformation correction is performed on the target image unit. The parameters of the geometric distortion degradation component describe the lens distortion coefficient or the projection deformation model caused by flight attitude. The grid-based elastic deformation correction achieves geometric correction by establishing a control point grid mapping relationship before and after deformation and resampling the image. In practical implementation, based on the influence model of the optical attenuation degradation component in the degradation feature descriptor, atmospheric light curtain estimation and contrast restoration operations are performed on the target image unit. Atmospheric light curtain estimation estimates the atmospheric illumination value by analyzing the brightest region or dark channel prior in the image. The contrast restoration operation compensates for the contrast decrease caused by atmospheric scattering by adjusting the dynamic range of the image pixel values. Optionally, based on the model and weights of various noise degradation components in the degradation feature descriptor, the target image unit is processed using a combination of adaptive filtering and wavelet domain thresholding. In practical implementation, the intermediate results after the correction and compensation operations performed on the motion blur degradation component, geometric distortion degradation component, optical attenuation degradation component, and various noise degradation components are fused. The fusion process requires allocating fusion coefficients according to the weights of different degradation components in the degradation feature descriptor. Refer to Table 1, which shows a specific example of the mapping relationship between degradation components, optimization operations, and parameter sources: Table 1: Mapping Table of Degraded Components and Optimization Processing Parameters In some embodiments, the intermediate results are fused using a pixel-level weighted fusion method. Each degradation component's corresponding correction and compensation operation generates an intermediate image. All intermediate images and the original target image units participate in the fusion process to generate the final optimized image unit. In a specific implementation, the fusion coefficients used for weighted fusion are dynamically calculated based on the influence weights of each degradation component in the degradation feature descriptor. The fusion process is implemented through a linear combination formula, expressed as follows: in: This represents the final optimized image unit. This represents the total number of degenerate component types participating in the fusion. Indicates assignment to the first Each degenerate component corresponds to a fusion coefficient in the optimization operation, where It preserves the original target image unit. Content integration coefficient. Represents the original target image unit Applying measures against the first The intermediate image obtained after correction and compensation operations on each degraded component. All fusion coefficients. The sum of is 1, and each The value of the degradation feature descriptor in the first The influence weights of the degenerate components are positively correlated; the degenerate component with a larger influence weight will have a correspondingly larger fusion coefficient in the optimization operation result. It can be understood that the multi-level guided focused retrieval process also includes a retrieval path backtracking verification step.

[0028] In specific implementations, after locking onto the target image units and forming a cluster of units to be processed, the system automatically performs reverse simulation verification along the recorded complete retrieval path from the entry node to each target image unit. Optionally, the simulation verification is performed by calculating whether the same target image units can still be stably retrieved along the same nodes and directed connection sequences, assuming the current degenerate feature descriptor is known. The stability criterion is that the combined score of the matching degree of each node and the cumulative strength of the path during the retrieval process is not lower than a certain percentage of the score at the time of the initial retrieval. In some embodiments, if the success rate of the simulation verification is lower than a preset value, the verified complete retrieval path is marked as an unstable path, and the dependency weight on this complete retrieval path or its topologically similar paths is reduced in subsequent retrieval processes. The reduction of dependency weight can be achieved by multiplying the cumulative strength of the path by a penalty factor less than 1.

[0029] See Figure 4This is a bar chart comparing the scores of drone aerial images before and after quality optimization, primarily used to quantitatively evaluate the actual effectiveness of the image optimization method. All images showed significant score improvements after optimization; the score improvement for a single image was between 20 and 25 points, indicating good adaptability of the optimization method to different image types. This chart serves to verify the effectiveness of the drone aerial image quality optimization method. The significant improvement in scores after optimization proves that the method can effectively solve degradation problems such as motion blur and noise; the stable optimization effect across different images indicates that the method is applicable to most aerial photography scenarios; the score of aerial image 4 is relatively low after optimization, allowing for a retrospective analysis of its degradation characteristics and targeted adjustment of optimization parameters.

[0030] In one embodiment of the present invention, the quality difference metric is obtained by calculating and weighting the differences between the optimized image unit and the original target image unit across multiple visual feature dimensions. Simultaneously, the absolute quality score of the optimized image unit is calculated, which is generated independently for the optimized image unit by a pre-trained full-reference or no-reference image quality assessment model. In some embodiments, based on the calculated quality difference metric and absolute quality score, a new node is created for the optimized image unit in the adaptive search structure, or the feature representation of the node corresponding to the original target image unit is updated. The decision is based on whether the quality difference metric exceeds a set update threshold. If it does, a new node is created to record this significant optimization trajectory; if it does not, the feature vector of the original node is updated to reflect its slight quality improvement. It can be understood that, based on the effectiveness of the optimization process and the effectiveness of the retrieval path, the strength of all connections on the path from the entry node to the new or updated node in the adaptive search structure is adjusted. The effectiveness of the optimization process is quantified by the increase in the absolute quality score, and the effectiveness of the retrieval path is quantified by the path stability score recorded in the previous retrieval path backtracking verification step. In practice, connections that lead to successful optimization are strengthened, while invalid or inefficient connections are weakened. The adjustment of connection strength is achieved through a feedback-based incremental learning mechanism, which uses the quality optimization result as a reward signal to update the weights of relevant connections. Optionally, the newly generated optimized image units and their metadata are archived to an aerial image library, and the spatial-quality association mapping table is updated synchronously. The archiving operation includes storing the image data of the optimized image units, the associated degradation feature descriptors, the optimization processing parameters used, and the calculated absolute quality score.

[0031] In some embodiments, the formula for calculating the quality difference metric is expressed as: in: This represents the final calculated measure of quality difference. This represents the total number of visual feature dimensions involved in the measurement. Indicates the first Preset weights for each visual feature dimension in difference fusion. Indicates the first In each visual feature dimension, the optimized image unit With the original target image unit The difference function. Visual feature dimensions may include, but are not limited to, edge sharpness, texture complexity, noise variance, or structural similarity index. The specific form depends on the nature of the selected feature dimension.

[0032] It is understandable that this is based on quality variation measurement. Compared with the preset update threshold The comparison results determine the execution of different node operations; this decision-making logic is implemented through a conditional judgment. In practical implementation, when quality difference measurement... Greater than or equal to the update threshold At that time, the system creates a new node in the adaptive search structure to represent the optimized image unit. The feature vector of the new node is obtained by extracting features from the optimized image unit. At the same time, directed connections are established from the corresponding node of the original target image unit to the new node, and the quality difference is measured. The strength of this directed connection is initialized with the absolute quality score. Optionally, when the quality difference metric... Less than the update threshold At this time, the system does not create new nodes, but updates the feature representation of the nodes corresponding to the original target image units. The update method is to perform a weighted average of the new feature vector calculated based on the optimized image units and the feature vector of the original nodes, with the weights measured by the quality difference. The decision is made. In practice, the adjustment of connection strength follows an update rule based on validity feedback. For each directed connection on the path from the entry node to the new or updated node, the adjustment amount of its strength is determined. The effectiveness of the optimization process and the effectiveness of the retrieval path are positively correlated, and this relationship can be expressed by a product factor model.

[0033] See Figure 5This is a quality score trend chart of drone aerial image optimization in stages, primarily used to demonstrate the progressive improvement effect of different optimization steps on image quality. With each optimization step, the image quality shows a stable upward trend without any reverse fluctuations. From the "original image" to the "denoising" stage, the score increases from approximately 0.45 to 0.65, representing the largest single-step improvement. The final "comprehensive optimization" stage achieves a score close to 0.9, representing the optimal result after all steps are combined. This chart serves to analyze the step-by-step effectiveness of the drone image optimization process. Denoising is a core preliminary step for improving image quality, and its parameter accuracy must be prioritized. The improvement in later steps is relatively gradual, allowing for a more balanced allocation of computational resources across steps.

[0034] 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.

[0035] 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 optimizing the image quality of UAV aerial photography areas based on image processing, characterized in that, include: Receive raw image streams and flight status data streams synchronously collected by the UAV platform during flight operations; Perceptual degradation analysis is performed on the original image stream to identify multiple degradation factors, including atmospheric disturbances, inherent noise of the imaging device, and geometric distortion caused by flight conditions. The degree of degradation caused by each degradation factor to the current image stream is quantified, and a degradation feature descriptor containing multiple degradation components and their weights is constructed. Based on the degradation feature descriptor, an adaptive search structure covering the target region and driven by image quality expectations is dynamically constructed. The adaptive search structure organizes the image units in the image library into a topological network that can reflect the quality evolution path. In the adaptive search structure, a multi-level guided focused retrieval is performed based on the degradation type and intensity indicated by the degradation feature descriptor, filtering and locking a series of target image units most relevant to the degradation feature descriptor layer by layer to form a cluster of units to be processed; The cluster of units to be processed is subjected to quality optimization processing. The quality optimization processing adopts a hybrid enhancement strategy based on the parameter adaptation of the degradation feature descriptor. Correction and compensation operations are applied to different degradation components to generate optimized image units. The optimized image units are then fed back into the adaptive search structure to update their network state.

2. The method for optimizing the image quality of UAV aerial photography areas based on image processing according to claim 1, characterized in that, The process involves perceptual degradation analysis of the original image stream, identifying multiple degradation factors including atmospheric disturbances, inherent noise from the imaging device, and geometric distortion caused by flight conditions. It also quantifies the degree of degradation each degradation factor causes to the current image stream, thereby constructing a degradation feature descriptor containing multiple degradation components and their weights, including: A continuous frame sequence is extracted from the original image stream. By analyzing the inter-frame differences and intra-frame statistical characteristics, dynamic degradation components and static degradation components are initially separated. The attitude angular velocity and position jitter information in the flight status data stream are spatiotemporally aligned and correlated with the dynamic degradation components to calculate the specific parameters of the three degradation factors caused by flight motion: motion blur, geometric distortion, and field of view instability. Multi-scale texture analysis and noise power spectrum estimation are performed on single-frame images in the original image stream to separate the influence models of three degradation factors: fixed pattern noise determined by the characteristics of the imaging sensor, random noise, and optical attenuation caused by atmospheric scattering. The preliminary results after correlation of flight status data stream and single-frame analysis are cross-validated to quantify the confidence and influence weight of each degradation factor. The parameters, influence models and their confidence and weight of all degradation factors are integrated into a structured data object, which is the degradation feature descriptor.

3. The method for optimizing the image quality of UAV aerial photography areas based on image processing according to claim 2, characterized in that, The method of dynamically constructing an adaptive search structure based on degraded feature descriptors, covering the target region and driven by image quality expectations, includes: The metadata and shallow visual features of all image units in the aerial image library are read in advance. Based on the acquisition time, geographical location and basic quality score of the image units, a preliminary spatial-quality correlation mapping table is established. Guided by the dominant degradation type in the degradation feature descriptor, image units that have been historically affected by similar degradation problems and have shown significant optimization effects are selected from the spatial-quality correlation mapping table and used as the initial seed nodes for constructing the adaptive search structure. Using the initial seed node as the core, directed connections are established between nodes based on the content coherence, quality inheritance relationship and spatiotemporal proximity between image units. The direction of the connection represents the direction that quality optimization may be passed or referenced, and the strength of the connection is determined by the similarity of features between nodes and the correlation of quality gradient. An online update mechanism is introduced, which adjusts the connection strength and topological relationship of nodes in the adaptive search structure in real time based on the latest received original image stream and its corresponding degradation feature descriptor. For newly emerging degradation patterns, the online update mechanism creates new node branches in the adaptive search structure to accommodate and represent the degradation patterns.

4. The method for optimizing the image quality of UAV aerial photography areas based on image processing according to claim 3, characterized in that, In the adaptive search structure, based on the degradation type and intensity indicated by the degradation feature descriptor, a multi-level guided focused retrieval is performed, progressively filtering and locking onto a series of target image units most relevant to the degradation feature descriptor, forming a cluster of units to be processed, including: The degenerate feature descriptor is encoded into a query vector that the adaptive search structure can understand, and the query vector is input into the entry node of the adaptive search structure; The first-level search is initiated. Within the direct neighborhood of the entry node, the matching degree between the query vector and the feature vector of the image unit represented by each neighboring node is calculated. Nodes with matching degrees exceeding the primary threshold are selected to form the primary candidate set. The second-level search is initiated, with each node in the primary candidate set as a new search starting point, and the search is explored downstream of the adaptive search structure. The exploration depth is limited by a preset number of hops. During the downstream exploration, not only is the direct matching degree between the query vector and the downstream node calculated, but the cumulative effect of the connection strength on the path from the primary candidate node to the downstream node is also evaluated. The scores of both are combined to select secondary candidate nodes. All secondary candidate nodes are deduplicated and merged, sorted according to their comprehensive scores, and the set of nodes that rank high and meet the final threshold is selected. The image unit corresponding to the set of nodes is locked as the target image unit. All target image units together constitute the unit cluster to be processed. At the same time, the complete retrieval path from the entry node to each target image unit is recorded.

5. The method for optimizing the image quality of UAV aerial photography areas based on image processing according to claim 4, characterized in that, The quality optimization process performed on the cluster of units to be processed employs a hybrid enhancement strategy based on parameter adaptation according to degenerate feature descriptors, including: The attributes of each target image unit in the cluster of units to be processed are analyzed, and the retrieval path information associated with the target image unit is extracted from the adaptive search structure; Based on the parameters for motion blur degradation components in the degradation feature descriptor, deconvolution filtering with specific direction and intensity is applied to the target image unit; based on the parameters for geometric distortion degradation components in the degradation feature descriptor, mesh-based elastic deformation correction is performed on the target image unit. Based on the influence model of optical attenuation degradation component in the degradation feature descriptor, atmospheric light curtain estimation and contrast restoration operations are performed on the target image unit; Based on the models and weights of various noise degradation components in the degradation feature descriptor, the target image unit is processed by a combination of adaptive filtering and wavelet domain threshold denoising. The intermediate results after the correction and compensation operations for motion blur, geometric distortion, optical attenuation and various types of noise are fused together. During the fusion process, the fusion coefficients are assigned according to the weights of different degradation components to generate the final optimized image unit.

6. The method for optimizing the image quality of a UAV aerial photography area based on image processing according to claim 5, characterized in that, The step of feeding the optimized image units back into the adaptive search structure to update their network state includes: Calculate the quality difference measure between the optimized image unit and the original target image unit in the unit cluster to be processed, and the absolute quality score of the optimized image unit; Based on the quality difference metric and absolute quality score, a new node is created for the optimized image unit in the adaptive search structure, or the feature representation of the node corresponding to the original target image unit is updated. Based on the effectiveness of the optimization process and the retrieval path, the strength of all connections on the path from the entry node to the new or updated node in the adaptive search structure is adjusted. Connections that lead to successful optimization are strengthened, while invalid or inefficient connections are weakened. The newly generated optimized image units and their metadata are archived into the aerial image library, and the spatial-quality association mapping table is updated synchronously.

7. The method for optimizing the image quality of UAV aerial photography areas based on image processing according to claim 5, characterized in that, The process involves processing the target image unit using a combination of adaptive filtering and wavelet domain thresholding, based on the models and weights for various noise degradation components in the degradation feature descriptor. Extract noise model parameters and their respective influence weights for fixed-pattern noise and random noise from the degradation feature descriptor; Based on the model parameters of the fixed-pattern noise, the spatial distribution pattern of the noise is located on the target image unit, and a matching spatial variation filter is constructed. The kernel function of the spatial variation filter is adaptively adjusted according to the local statistical characteristics of the noise pattern to suppress the fixed-pattern noise. Based on the model parameters of the random noise, its noise power spectrum characteristics are determined, and the optimal number of decomposition layers for wavelet transform is calculated, as well as different threshold functions are set for different sub-bands. Wavelet transform is performed on the image after spatial variation filtering to obtain high-frequency subband coefficients at multiple scales and directions; Based on the threshold function, each high-frequency subband coefficient is subjected to soft or hard thresholding. Coefficients with amplitudes below the threshold are treated as noise and suppressed, while coefficients with amplitudes above the threshold are retained to maintain image details. The high-frequency subband coefficients after thresholding are subjected to inverse wavelet transform to reconstruct the image after removing random noise. Based on the influence weights of fixed-pattern noise and random noise extracted from the degraded feature descriptor, the output of spatial variation filtering and the output of wavelet threshold denoising are weighted and fused to generate the final noise-suppressed image.

8. The method for optimizing the image quality of UAV aerial photography areas based on image processing according to claim 2, characterized in that, The perceptual degradation analysis also includes an environmental context modeling step, including: Collect flight environment data synchronized with the original image stream, the flight environment data including at least light intensity, atmospheric visibility and temperature and humidity; An empirical correlation model between environmental data and image degradation factors is established, and the degradation factor parameters calculated by image content analysis and flight status correlation are calibrated using the empirical correlation model for environmental context calibration. The calibrated parameters are used as more accurate inputs to update and correct degenerate feature descriptors.

9. The method for optimizing the image quality of UAV aerial photography areas based on image processing according to claim 3, characterized in that, The dynamic construction process of the adaptive search structure also includes a structure self-optimization step, including: Regularly evaluate the retrieval efficiency and accuracy of the adaptive search structure in historical queries; When the retrieval efficiency or accuracy is lower than the set standard, the structure self-optimization step is triggered. The structure self-optimization step iteratively adjusts the node division rules, connection establishment criteria, and connection strength calculation formula of the adaptive search structure based on historical query logs and optimization result feedback. While ensuring compatibility with existing data, the adjusted new rules are applied to the adaptive search structure to complete the reconstruction of its topology and the reconfiguration of its parameters.

10. The method for optimizing the image quality of UAV aerial photography areas based on image processing according to claim 4, characterized in that, The multi-level guided focused retrieval process also includes a retrieval path backtracking verification step, including: After locking onto the target image unit and forming a cluster of units to be processed, the system automatically performs reverse simulation verification along the recorded complete retrieval path; Simulation verification is performed by calculating whether the same target image unit can still be stably retrieved along the same path, assuming that the current degraded feature descriptor is known. If the success rate of the simulation verification is lower than the preset value, the search path will be marked as an unstable path, and the dependence weight on the search path or its similar paths will be reduced in subsequent searches.

Images