Unmanned aerial vehicle image noise reduction and color enhancement optimization method and system

Abstract

The invention relates to the technical field of unmanned aerial vehicle image processing, and discloses an unmanned aerial vehicle image noise reduction and color enhancement optimization method and system. The method comprises the following steps: firstly, acquiring an original image sequence acquired in the flight process of the unmanned aerial vehicle, and identifying corresponding environment illumination intensity change data and motion fuzzy feature distribution; time dimension features of the original image sequence are divided based on environment illumination intensity change data, and a noise feature distribution map is constructed in combination with motion blur feature distribution; and finally, calling an equipment imaging parameter set of the original image sequence, analyzing a corresponding imaging mode constraint condition, and formulating a multi-stage noise reduction strategy framework for the original image sequence according to the constraint condition and the noise feature distribution map. The method realizes effective noise reduction and color enhancement of the unmanned aerial vehicle image by comprehensively considering the environment illumination change, the motion blur feature and the equipment imaging parameter, improves the image quality, and is suitable for the field of unmanned aerial vehicle image processing.

Description

Technical Field

[0001] This invention relates to the field of UAV image processing technology, specifically to a method and system for UAV image noise reduction and color enhancement optimization. Background Technology

[0002] With the rapid development of drone technology, its applications in aerial surveying, environmental monitoring, agricultural plant protection, emergency rescue, and many other fields are becoming increasingly widespread. Images captured by drones serve as a crucial carrier of information, and their quality directly impacts the effectiveness of subsequent data analysis and decision-making. However, during actual flight, drones are often affected by complex environmental factors and the inherent characteristics of the equipment itself, resulting in various quality issues in the raw images they acquire. Ambient lighting conditions are one of the key factors affecting image quality. When flying, drones may face drastic changes in light during sunrise and sunset, fluctuating light levels due to cloud cover, or uneven light intensity in different areas. These dynamic changes in lighting can cause problems such as color shifts and decreased contrast in images. Especially in low-light environments, image noise is more pronounced, severely impacting the visual quality and the accuracy of information extraction.

[0003] The motion of a drone can also significantly impact image quality. Due to the drone's own vibrations, airflow disturbances, and rapid adjustments in flight attitude, motion blur can easily occur in the acquired images. This blur not only reduces image sharpness but also causes the loss of detailed information, posing a significant challenge to subsequent image analysis and processing.

[0004] Currently, there has been some research on noise reduction and enhancement methods for UAV images, but most existing methods have limitations. Some methods only address image noise or a single type of blur under specific environments, lacking adaptability to dynamically changing environments; others do not fully consider the parameter characteristics of UAV imaging equipment, resulting in processing effects that do not match actual imaging needs and fail to meet the requirements of high-precision image processing. Therefore, developing an image noise reduction and color enhancement optimization method that can comprehensively address changes in ambient lighting and motion blur, and adapt to the imaging parameters of the equipment, has become a problem that needs to be solved in the field of UAV image processing. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for denoising and enhancing the color of drone images, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for denoising and enhancing the color of UAV images, the method comprising: Acquire the original image sequence collected during the flight of the UAV, and identify the ambient light intensity change data and motion blur feature distribution corresponding to the original image sequence; Based on the ambient light intensity change data, the temporal dimension features of the original image sequence are divided, and combined with the motion blur feature distribution, a noise feature distribution map of the original image sequence is constructed. The device imaging parameter set of the original image sequence is invoked, the imaging mode constraints corresponding to the device imaging parameter set are parsed, and a multi-level noise reduction strategy framework for the original image sequence is formulated based on the imaging mode constraints and the noise feature distribution map.

[0007] Preferably, the step of formulating a multi-level noise reduction strategy framework for the original image sequence based on the imaging mode constraints and the noise feature distribution map includes: Extract the high-frequency and low-frequency noise components from the noise feature distribution map, and calculate the energy distribution ratio of the high-frequency and low-frequency noise components in the time dimension; correlate the exposure time threshold and the photosensitive element sensitivity parameter in the imaging mode constraints to determine the suppression intensity level corresponding to the high-frequency noise component and the smoothing depth corresponding to the low-frequency noise component; based on the suppression intensity level and the smoothing depth, generate the spatial domain filtering rules and temporal domain filtering rules in the multi-level noise reduction strategy framework.

[0008] Preferably, after constructing the noise feature distribution map of the original image sequence, the method further includes: The color space conversion data of the original image sequence is collected, and the color gamut coverage and color distortion probability distribution corresponding to the color space conversion data are analyzed. Based on the color gamut coverage, a reference color mapping relationship of the original image sequence is established. Combined with the color distortion probability distribution, the correctable parameter range of the reference color mapping relationship is calculated. Based on the correctable parameter range, a color enhancement constraint boundary for the original image sequence is generated.

[0009] Preferably, after generating the color enhancement constraint boundary for the original image sequence, the method further includes: The multi-level noise reduction strategy framework is executed to process the original image sequence to obtain a set of intermediate images after noise reduction. The local contrast features and global average brightness of the intermediate image set are extracted and associated with the upper limit threshold of saturation and the hue balance coefficient in the color enhancement constraint boundary. Based on the local contrast features and the global average brightness, the applicable area range of the upper limit threshold of saturation in the intermediate image set is dynamically adjusted, and the weight distribution ratio of the hue balance coefficient in the shadow area and the highlight area is simultaneously corrected.

[0010] Preferably, the dynamic adjustment of the applicable region range of the upper limit saturation threshold in the intermediate image set includes: The intermediate image set is divided into landform type regions and artificial building regions. The natural color distribution probability of the landform type region and the texture complexity index of the artificial building region are calculated. The vegetation cover sensitivity coefficient and water reflection feature parameters in the applicable area are correlated to establish a color correction priority sequence between the natural color distribution probability and the texture complexity index. Based on the color correction priority sequence, the first gradient adjustment step size of the saturation upper limit threshold in the landform type region and the second gradient adjustment step size in the artificial building region are configured.

[0011] Preferably, after formulating the multi-level noise reduction strategy framework for the original image sequence, the method further includes: The system calls upon a reference image database from historical flight missions and matches the scene similarity probability distribution between the reference image database and the current original image sequence. It then analyzes the matching degree of illumination conditions and the overlap degree of ground feature contours in the scene similarity probability distribution to generate a noise suppression reference template based on the reference image database. Finally, it fuses the noise suppression reference template with the spatial filtering parameters in the multi-level noise reduction strategy framework to optimize the motion blur compensation intensity in the temporal filtering rules.

[0012] Preferably, the synchronous correction of the weight distribution ratio of the tone balance coefficient in the shadow and highlight areas includes: The boundary coordinates of the shadow region and the proportion of overexposed pixels in the highlight region are detected in the intermediate image set. The geometric continuity index of the boundary coordinates of the shadow region and the halo diffusion range of the overexposed pixel ratio are calculated. The shadow detail preservation factor and the highlight suppression attenuation parameter in the weight allocation ratio are associated to construct a tone correction linkage mechanism between the geometric continuity index and the halo diffusion range. Based on the tone correction linkage mechanism, the optimal tone compensation curve of the shadow region and the optimal brightness attenuation curve of the highlight region are output.

[0013] Preferably, after configuring the upper limit threshold for saturation in the first gradient adjustment step size of the landform type region, the method further includes: Identify the vegetation cluster distribution characteristics and soil bare area range in the landform region, measure the chlorophyll reflectance intensity of the vegetation cluster distribution characteristics and the infrared absorptivity of the soil bare area range; correlate the vegetation health correction coefficient and soil moisture influence factor in the first gradient adjustment step size, and establish a saturation adaptive adjustment model between the chlorophyll reflectance intensity and the infrared absorptivity; dynamically control the execution frequency of the first gradient adjustment step size in dense vegetation areas and soil bare areas through the saturation adaptive adjustment model.

[0014] Preferably, after obtaining the set of intermediate images after noise reduction, the method further includes: Verify the signal-to-noise ratio improvement and detail retention score of the intermediate image set. When the signal-to-noise ratio improvement is lower than a preset threshold, backtrack the spatial domain filtering rule parameters in the multi-level noise reduction strategy framework. Simultaneously detect the color consistency distribution and the smoothness of the transition between adjacent frames of the intermediate image set. When the color consistency distribution deviates from the target color temperature range, recalibrate the white balance reference point in the color enhancement constraint boundary.

[0015] Preferably, the present invention also includes a UAV image denoising and color enhancement optimization system for implementing the UAV image denoising and color enhancement optimization method described above, the system comprising: The image feature analysis module is used to acquire the original image sequence collected during the flight of the UAV, identify the ambient light intensity change data and motion blur feature distribution corresponding to the original image sequence, divide the time dimension features based on the ambient light intensity change data, and construct a noise feature distribution map by combining the motion blur feature distribution. The noise reduction strategy generation module is used to call the device imaging parameter set and parse the imaging mode constraints, and formulate a multi-level noise reduction strategy framework based on the imaging mode constraints and the noise feature distribution map. The color constraint construction module is used to collect color space conversion data, analyze the color gamut coverage and color distortion probability distribution, establish a baseline color mapping relationship and calculate the correctable parameter range, and generate color enhancement constraint boundaries. The collaborative optimization execution module is used to execute a multi-level noise reduction strategy framework to generate an intermediate image set, extract local contrast features and global brightness mean, dynamically adjust the applicable area range of the upper limit threshold of saturation and correct the weight allocation ratio of the tone balance coefficient. The output verification module is used to verify the signal-to-noise ratio improvement and detail retention score, backtrack the spatial domain filtering rule parameters, detect the color consistency distribution and the smoothness of the transition between adjacent frames, and calibrate the white balance reference point.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By acquiring ambient light intensity variation data and motion blur feature distribution corresponding to the original image sequence, key factors affecting image quality can be comprehensively captured. Dynamic changes in ambient light intensity and the feature distribution of motion blur are the main causes of image quality degradation in UAVs. Accurate identification of these two factors provides a targeted basis for subsequent processing.

[0017] By dividing the original image sequence into temporal features based on ambient light intensity variation data and constructing a noise feature distribution map by combining motion blur feature distribution, a deeper and more systematic understanding of image noise is achieved. The division of temporal features reflects the changing patterns of images under different lighting conditions at different times, while the noise feature distribution map clearly presents the distribution of noise in the image, avoiding the blindness and one-sidedness of traditional methods in noise processing.

[0018] By invoking the device's imaging parameter set and parsing the corresponding imaging mode constraints, a multi-level noise reduction strategy framework is formulated based on these constraints and the noise feature distribution map, enhancing the method's practicality and adaptability. Different imaging devices have different parameter characteristics; the introduction of imaging mode constraints ensures that the noise reduction strategy matches the actual imaging capabilities of the device, avoiding poor processing results caused by deviating from device characteristics. The multi-level noise reduction strategy framework can take corresponding processing measures according to different noise characteristics and distributions, achieving effective processing of different types and levels of noise, while also considering color enhancement. This results in improved image sharpness and color performance, better meeting the image quality requirements of practical application scenarios. Attached Figure Description

[0019] Figure 1 This is a timing diagram of the UAV image denoising and color enhancement optimization method described in this invention; Figure 2 A flowchart outlining the multi-level noise reduction strategy framework; Figure 3 A flowchart for generating color-enhanced constraint boundaries; Figure 4 A flowchart illustrating the coordinated execution of noise reduction and color enhancement; Figure 5 This is a flowchart for dynamically adjusting the saturation threshold. Detailed Implementation

[0020] 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.

[0021] Please see Figure 1 This invention provides a method for denoising and enhancing the color of drone images, the method comprising: By acquiring raw image sequences during UAV flight, this method identifies ambient light intensity variation data and motion blur feature distribution. Temporal features are segmented based on the ambient light intensity variation data, and a noise feature distribution map is constructed by combining the motion blur feature distribution. The device's imaging parameter set is used to analyze imaging mode constraints, and a multi-level noise reduction strategy framework is formulated based on these constraints and the noise feature distribution map. This method achieves adaptive noise reduction by establishing a correlation model between noise features and imaging parameters. Simultaneously, by combining the analysis of color space conversion data, color enhancement constraint boundaries are constructed, ultimately outputting high-quality images.

[0022] Example 1: See Figure 2 High-frequency and low-frequency noise components are extracted from the noise feature distribution map. Based on the energy distribution ratio in the time dimension and combined with the imaging mode constraints, spatial domain filtering rules and temporal domain filtering rules are formulated in the multi-level noise reduction strategy framework.

[0023] High-frequency noise components mainly manifest as random particle noise in the image, typically caused by sensor noise, electronic interference, or quantization errors. This type of noise is concentrated in the high-frequency range, appearing as small, irregular particles in image detail areas. Low-frequency noise components, on the other hand, manifest as large-area color block noise, usually caused by uneven illumination, sensor dark current, or thermal noise from long-term exposure. This noise is concentrated in the low-frequency range, appearing as slowly changing color blocks in smooth areas of the image. The noise feature distribution map is constructed by analyzing the time-frequency characteristics of the image sequence. A multi-scale decomposition method is used to separate noise components in different frequency bands, and their distribution patterns along the time axis are statistically analyzed.

[0024] To calculate the energy distribution ratio of high-frequency noise components to low-frequency noise components over time, a sliding window analysis technique is employed to perform frame-by-frame spectral analysis of the image sequence. The energy proportion of high-frequency noise is determined by calculating the average power spectral density of the high-frequency components, while the energy proportion of low-frequency noise is measured by the integrated energy of the low-frequency components. The temporal trend is obtained by calculating the noise energy difference between adjacent frames, forming a curve model of noise energy changing over time. This model can reflect the impact of factors such as changes in ambient lighting and adjustments in motion state during UAV flight on noise distribution.

[0025] The exposure time threshold and sensor sensitivity parameters in the imaging mode constraints directly affect noise characteristics. When the exposure time exceeds the set threshold, thermal noise accumulation leads to a significant increase in low-frequency noise energy; when the sensitivity exceeds a critical value, the sensor gain amplifies high-frequency noise components. By analyzing the device's imaging parameter set, a mapping relationship between exposure time, sensitivity, and noise energy is established. The suppression intensity level corresponding to high-frequency noise components is dynamically adjusted according to their energy proportion; the higher the energy proportion, the greater the suppression intensity. The smoothing depth corresponding to low-frequency noise components has a non-linear relationship with their energy distribution; light smoothing is used when the energy is low, and deep smoothing is used when the energy is high to eliminate large-area color block noise.

[0026] The spatial domain filtering rules are generated based on the spatial distribution characteristics of noise components. High-frequency noise suppression employs an adaptive filtering method, dynamically selecting the filter type and parameters according to the noise intensity of local regions. For textured regions, an edge-protection filter is used to avoid detail loss; for flat regions, a strong noise suppression filter is used to improve the signal-to-noise ratio. Low-frequency noise smoothing uses a non-local mean algorithm, which performs weighted averaging by searching for similar regions in the image, effectively eliminating large-area color block noise while preserving image structural information.

[0027] The temporal filtering rules are formulated considering the temporal correlation of the image sequence. Motion blur feature distribution is used to compensate for the blurring effect caused by inter-frame motion. A temporal recursive filtering model is established by estimating the motion vector field between adjacent frames. High-frequency noise exhibits strong randomness in the temporal domain, so short-term memory filtering is used to suppress instantaneous noise; low-frequency noise has strong correlation in the temporal domain, so long-term memory filtering is used to smooth slowly changing noise components. The motion blur compensation intensity is adaptively adjusted according to the UAV's flight speed and attitude change rate, increasing the compensation intensity during high-speed motion and decreasing it during low-speed motion to avoid over-smoothing.

[0028] The final output of the multi-level noise reduction strategy framework includes a spatial domain filtering parameter matrix and temporal domain filtering weight coefficients. The spatial domain parameter matrix records the filter combinations and their parameter configurations corresponding to different noise levels, while the temporal weight coefficients define the strength and range of inter-frame filtering. This framework can automatically select the optimal noise reduction strategy based on real-time acquired image features and imaging parameters, achieving efficient noise suppression for UAV images.

[0029] In the specific implementation, the suppression intensity level of high-frequency noise components is controlled by a graded threshold, dividing the noise energy range into multiple intervals, each corresponding to a different filtering intensity. The smoothing depth of low-frequency noise components adopts a multi-scale analysis method, selecting smoothing kernels of different scales based on the size and intensity of the noise block. The execution order of spatial domain filtering rules follows a progressive processing flow from high frequency to low frequency, first suppressing fine particle noise, and then smoothing large-area color block noise to avoid mutual interference between different noise components.

[0030] The optimization of temporal filtering rules considers a balance between computational efficiency and processing effect. The frame range of recursive filtering is dynamically adjusted according to computing resources, utilizing as much frame information as possible while ensuring real-time performance. Motion blur compensation adopts a pyramid hierarchical matching algorithm, first estimating coarse motion vectors at a large scale, and then gradually refining them to a small scale to improve the accuracy and robustness of motion estimation. The final noise reduction strategy framework can adapt to the imaging needs of UAVs in complex environments and effectively improve image quality.

[0031] Example 2: See Figure 3 The color space conversion data of the original image sequence is collected and analyzed to establish a baseline color mapping relationship and generate color enhancement constraint boundaries. At the same time, the spatial and temporal filtering parameters in the noise reduction strategy framework are optimized by combining the reference image database from historical flight missions.

[0032] In processing color space conversion data, the conversion matrix from the device-dependent RGB color space to the standard Lab color space is first analyzed. This conversion process needs to consider the characteristics of the UAV's imaging sensor, including the spectral response curve, white balance preset value, and the influence of color interpolation algorithms. Color gamut coverage is quantified by comparing the volume difference between the device's RGB space and the standard color gamut space, and the degree of overlap between the two on the CIExy chromaticity map is calculated. The probability distribution of color distortion is obtained by statistically analyzing the offset of pixel values ​​in each channel before and after conversion, and a color difference histogram model is established to analyze typical distortion patterns. Common color distortions include hue shift in highlight areas, saturation loss in shadow areas, and nonlinear response of specific color channels.

[0033] The baseline color mapping relationship is established using a multinomial regression method, mapping the color values ​​captured by the device to a standard color space. The mapping process preserves the reversible transformation properties, ensuring that the enhanced image accurately reproduces the colors of the original scene. The correctable parameter range is determined by analyzing the standard deviation of the color difference distribution; abnormal color values ​​exceeding the 3σ range are marked as uncorrectable regions. These regions typically correspond to sensor saturation, severe noise interference, or extreme colors exceeding the device's color gamut. The generation of color enhancement constraint boundaries considers color gamut limits and visual perception characteristics, defining an upper limit for saturation adjustment and a tolerance range for hue change in the CIELab color space to avoid over-enhancement leading to color distortion.

[0034] The reference image database from historical flight missions contains typical scene images under different lighting conditions, seasonal variations, and geographical features. Scene similarity analysis between the reference images and the current original image sequence employs deep feature matching technology, extracting high-level semantic features through a pre-trained convolutional neural network and calculating the cosine similarity score between feature vectors. Lighting condition matching is obtained by comparing the lighting distribution histograms of the images, analyzing the similarity in brightness, contrast, and color temperature. Ground feature contour overlap is determined using multi-scale structural similarity analysis, measuring the matching degree of image content at three levels: edge, texture, and global structure.

[0035] The noise suppression reference template is generated using a weighted fusion strategy, adaptively combining denoising parameters in the reference image based on scene similarity scores. The template contains optimal spatial filtering parameters for different scene types, such as Gaussian kernel size, chromatic aberration threshold for bilateral filtering, and nonlocal mean search window size. The optimization of the temporal filtering rules focuses on adjusting the motion blur compensation intensity. The reference template stores inter-frame displacement vector fields under typical motion patterns to guide motion estimation for the current image sequence. Dynamic adjustment of the spatial filtering parameters is based on the difference analysis between the noise feature distribution and the reference template. When the noise characteristics of the current image differ significantly from the reference scene, incremental learning is used to update the filtering parameters.

[0036] The synergistic optimization of color enhancement constraints and denoising strategies is reflected in the rational arrangement of processing order. Multi-level denoising is performed first to eliminate image noise, followed by color space conversion and enhancement operations, preventing noise from being abnormally amplified during color conversion. When performing enhancement processing in the Lab color space, the luminance channel (L) and color channels (ab) are processed separately to prevent brightness adjustments from affecting color balance. Typical color distributions from a reference image database are used to calibrate enhancement parameters, especially when dealing with extreme lighting conditions such as sunsets and snowy landscapes; the reference color features effectively prevent color shifts.

[0037] A real-time update mechanism for scene similarity probability distribution ensures the system can adapt to continuously changing shooting environments. After processing each image sequence segment, the feature vector of the current frame is compared with the reference database. When a new scene pattern is discovered, its features and optimal processing parameters are added to the database. The calculation of lighting condition matching degree adopts an adaptive threshold strategy, dynamically adjusting the matching sensitivity according to the rate of change of ambient lighting. Geometric consistency verification is introduced into the ground feature contour overlap analysis to eliminate false matches caused by brief occlusion or instantaneous movement.

[0038] The fusion of the noise suppression reference template and the multi-level noise reduction strategy framework employs a parametric interpolation method. For spatial filtering parameters, the most similar scene samples are found in the reference template, and linear interpolation is performed based on similarity weights. The motion blur compensation intensity in the temporal filtering rules is adjusted according to the real-time motion state of the UAV, and the accuracy of motion estimation is improved by combining inertial measurement unit data. The dynamic adjustment of the color enhancement constraint boundary considers the semantic information of the scene content, and differentiated enhancement strategies are adopted for different types of terrain features, such as emphasizing the protection of green hues in vegetation areas and maintaining neutral gray balance in building areas.

[0039] In terms of implementation details, color space conversion employs lookup table acceleration technology, pre-calculating the conversion matrix from the device's RGB to the standard Lab color space and storing it in memory. Color gamut boundary detection uses a convex hull algorithm to determine the limits of the device's color space, constructing polygonal boundaries on the chromaticity map. The historical reference image database adopts a hierarchical storage structure, establishing multi-level indexes according to lighting conditions, season type, and geographical features to improve scene retrieval efficiency. Noise suppression reference template updates utilize a sliding window mechanism, retaining frequently used templates and discarding outdated processing methods.

[0040] The color enhancement process incorporates a human visual characteristic model, applying enhancement levels to different color channels that align with visual sensitivity. In the Lab color space, the enhancement coefficients of the a / b channels exhibit a non-linear relationship with color saturation, preventing color noise in low-saturation regions. The reversibility of the baseline color mapping relationship is verified through round-trip conversion tests, ensuring that color enhancement does not introduce irreversible distortion. Maintenance of the reference image database includes regularly cleaning low-quality samples and supplementing with typical scenes to maintain the database's timeliness and representativeness.

[0041] The collaborative processing architecture for noise reduction and color enhancement adopts a pipelined design, with each processing module exchanging data through standardized interfaces. The spatial filtering module outputs intermediate results to the color analysis module, while the motion estimation information from the temporal filtering module is shared with the color enhancement module for reference. During system runtime, the processing depth is dynamically adjusted based on computing resources, simplifying some processing steps when real-time requirements are high and enabling full-process optimization when high image quality requirements are high. The final generated color enhancement constraint boundaries and noise reduction strategy parameter set can adapt to the image processing needs of UAVs in complex and ever-changing environments.

[0042] Example 3: See Figure 4 This method performs color enhancement and optimization on the intermediate image set after noise reduction. By extracting local contrast features and global average brightness, it dynamically adjusts the upper limit threshold of saturation and the tone balance coefficient, and establishes a linkage correction mechanism for shadow and highlight areas. This approach achieves adaptive color optimization for different lighting conditions while maintaining the natural look of the image.

[0043] The local contrast feature analysis of the intermediate image set employs a multi-scale gradient calculation method to detect the sharpness of image details at different resolutions. The global brightness mean calculation excludes the influence of overexposed and underexposed areas, only calculating the pixel average within the effective brightness range. The adjustment of the saturation upper limit threshold is based on the HSV color space model, applying non-linear stretching to the saturation component while maintaining hue. The optimization of the tone balance coefficient considers the dynamic range distribution characteristics of the image; a compensation curve with progressive characteristics is used in shadow areas, while a smoothly transitioning attenuation function is used in highlight areas.

[0044] Shadow region boundary detection employs an improved Canny operator, enhancing edge continuity through adaptive dual-threshold settings. Overexposed pixels in highlight areas are identified using a region-growing-based clustering method to avoid misclassifying isolated noise points as overexposed. Geometric continuity is calculated by measuring the rate of curvature change of boundary line segments, reflecting the naturalness of shadow boundaries. The quantification of halo diffusion range utilizes morphological processing to measure the extent of overexposed regions' expansion within their neighborhood.

[0045] The shadow detail preservation factor is dynamically adjusted based on the texture complexity within the area, with higher preservation weights assigned to areas with rich texture. The highlight suppression attenuation parameter is positively correlated with the degree of overexposure, with stronger attenuation applied to severely overexposed areas. The establishment of the tone correction linkage mechanism relies on the statistical relationship between shadow and highlight areas; a collaborative adjustment model is constructed by analyzing the correlation between the two in color distribution.

[0046] The optimal tone compensation curve for shadow areas is fitted using a piecewise function, improving the visibility of details in dark areas while maintaining hue accuracy. The mathematical expression of this curve is:

[0047] in, This represents the original hue value of the shadow area. The compensated hue value, , , These are the slope coefficients for each segment. , , For the intercept term, and The threshold points are segmented. The curve uses a quadratic function to enhance low-brightness areas, a logarithmic transformation to maintain a smooth transition in medium-brightness areas, and linear processing to avoid overcompensation in high-brightness areas.

[0048] The optimal brightness decay curve for highlight areas employs an exponential decay model, adaptively adjusting the decay rate based on overexposure levels. Curve parameters are correlated with the overall image exposure level, avoiding brightness discontinuities caused by local adjustments. Color consistency detection is achieved by analyzing the color difference distribution between adjacent frames in the Lab color space, quantifying the smoothness of color changes between frames.

[0049] The white balance reference point is calibrated using an automatic selection strategy based on scene content, prioritizing neutral color regions in the image as references. The color temperature range is set considering typical lighting conditions in the shooting environment, establishing a mapping relationship between different scene types and ideal color temperature values. When color consistency deviates from the target range, the key control points within the color enhancement constraint boundaries are recalculated.

[0050] Local contrast features are extracted using a joint spatial-frequency analysis method, calculating gradient magnitude in the spatial domain and analyzing energy distribution in the frequency domain. The global brightness mean is calculated using a weighted average method, assigning different weights based on regional importance. The dynamic adjustment of the saturation upper limit threshold considers the semantic meaning of regional content, employing differentiated restriction strategies for specific objects such as vegetation and sky.

[0051] The weighting of the tone balance coefficient incorporates a visual saliency model, granting higher color fidelity to areas of high observer attention. The coordinated correction of shadows and highlights establishes a feedback adjustment mechanism; when one adjustment exceeds a threshold, the other is automatically triggered for compensation. This collaborative optimization effectively avoids color banding or brightness jumps caused by processing only a single area.

[0052] The image quality assessment module continuously monitors the processing effect. When it detects a color shift or detail loss exceeding the tolerance limit, it automatically reverts to the previous processing stage to readjust parameters. The collaborative workflow of noise reduction and color enhancement employs an iterative optimization strategy to pursue the best visual effect while ensuring processing efficiency. The final output image maintains a natural look while achieving a balance between enhancing shadow details and suppressing highlights.

[0053] Example 4: See Figure 5 This method divides the intermediate image set into landform type regions and artificial building regions, and configures differentiated saturation adjustment strategies based on the characteristics of different regions. By analyzing the distribution characteristics of vegetation clusters and the infrared characteristics of bare soil areas, this method establishes a color correction priority sequence to achieve adaptive color enhancement for different land cover types.

[0054] In farmland monitoring scenarios captured by drones, the system first performs semantic segmentation on the images, dividing typical landform areas into three categories: healthy vegetation areas, sparse vegetation areas, and bare soil areas. Healthy vegetation areas are identified by an NDVI index greater than 0.6, sparse vegetation areas correspond to an NDVI index between 0.3 and 0.6, and bare soil areas are those with an NDVI index below 0.3. The identification of man-made building areas uses edge detection combined with shape analysis, focusing on extracting structures with regular geometric features such as roofs and roads.

[0055] Table 1: Results of Color Feature Analysis for Typical Landform Areas Region Type Average saturation Hue main peak Texture complexity Adjust priority Healthy vegetation 0.45 120° 0.72 1 Sparse vegetation 0.38 115° 0.65 2 bare soil 0.28 35° 0.54 3 artificial buildings 0.31 0° 0.81 4 The vegetation health correction coefficient was calculated based on chlorophyll reflectance spectrum characteristics, determined by analyzing the reflectance ratio of red light to near-infrared bands in multispectral images. Soil moisture influencing factors were estimated using radiation temperature data in the thermal infrared band; wetter soil areas exhibited lower surface temperatures and higher heat capacity. The first gradient adjustment step size for densely vegetated areas was set to a smaller value to avoid over-enhancement leading to color distortion; the second gradient adjustment step size for bare soil areas was relatively larger to compensate for the typically flat color issues in these areas.

[0056] In a specific case study of farmland image processing, the system detected a cornfield containing three typical regions: a well-grown central area (healthy vegetation), a marginal area with sparse vegetation (sparse vegetation), and bare areas around the field ridges. The chlorophyll reflectance spectrum intensity of the healthy vegetation area showed typical vegetation characteristic curves, with a distinct reflectance peak in the near-infrared band. The reflectance spectrum curve of the sparse vegetation area showed a transitional shape, while the bare soil area exhibited typical soil spectral characteristics, with strong reflectance in the visible light band.

[0057] Water body identification is achieved by analyzing the reflection characteristics of specific wavelengths. Clean water exhibits strong absorption characteristics in the near-infrared band. When a water body is detected, the system automatically reduces the saturation adjustment intensity to maintain the natural blue hue of the water. The texture complexity of building areas is determined by analyzing the contrast characteristics of the gray-level co-occurrence matrix. Areas with higher complexity employ a conservative saturation enhancement strategy to avoid amplifying noise on building surfaces.

[0058] The establishment of the color correction priority sequence comprehensively considers multiple factors: the health of vegetation areas, soil moisture, water purity, and the complexity of building structures. During processing, the system dynamically monitors color change trends in each area. When the saturation of a certain area approaches the preset upper limit, it automatically reduces the adjustment intensity for that area. This adaptive mechanism effectively prevents over-processing during color enhancement.

[0059] The frequency control for densely vegetated areas and bare soil areas employs a time-weighted algorithm, giving higher priority to areas with good recent processing results. The system maintains a historical database of regional characteristics, recording the response characteristics of each area in each processing iteration, providing a reference for current parameter adjustments. For example, a farmland area that repeatedly experiences saturation overflow will be marked as a sensitive area, and the adjustment magnitude will be automatically reduced in subsequent processing.

[0060] Special treatment for artificial building areas takes into account the characteristics of building materials. Concrete structures maintain a neutral gray balance, while metal roofs have their color contrast appropriately enhanced but saturation increases controlled. The system establishes a mapping relationship between material type and processing parameters by analyzing the typical reflective properties of building materials. For building areas that cannot be clearly identified, a conservative default processing scheme is adopted to ensure that no significant color deviation is introduced.

[0061] In actual urban aerial image processing, the system successfully distinguished different areas such as park green spaces, sports field lawns, asphalt roads, and building complexes. Park green spaces were treated with a more aggressive color enhancement strategy to highlight the vitality of the vegetation; sports field lawns maintained moderate saturation to avoid unnatural bright green hues on artificial turf; road areas largely retained their original color tones with only slight contrast optimization; building complexes were differentiated based on the characteristics of their exterior wall materials, with brick-red walls appropriately enhanced and gray concrete walls kept neutral.

[0062] The system continuously learns from processing experiences in different scenarios, constantly optimizing its region division rules and parameter adjustment strategies. After each processing cycle, operators can fine-tune the automated results; these manual adjustments are recorded and used to improve subsequent automated processing algorithms. This interactive learning mechanism enables the system to gradually adapt to various unique aerial photography scenarios, including seasonal vegetation changes, diverse architectural styles, and complex lighting conditions.

[0063] Example 5: The intermediate image set after noise reduction and color enhancement is subjected to quality verification and parameter optimization. The processing flow is dynamically adjusted through multi-dimensional evaluation indicators to ensure that the output image meets the requirements of subsequent applications. This method establishes a closed-loop feedback mechanism that can backtrack and optimize key processing parameters based on real-time quality detection results, achieving adaptive adjustment of the processing flow.

[0064] The image quality verification phase first quantifies the signal-to-noise ratio (SNR) improvement using a noise power spectrum comparison method based on frequency domain analysis. Images before and after denoising are converted to the frequency domain, and the change in noise energy in each frequency band is calculated, with a focus on noise suppression in the mid-to-high frequency regions. Detail retention is scored using multi-scale structural similarity analysis, comparing edge sharpness and texture features of images before and after processing at different resolutions. When the SNR improvement falls below a preset threshold, the system automatically triggers a denoising parameter backtracking mechanism, focusing on checking the core parameter settings in the spatial domain filtering rules.

[0065] The evaluation of color consistency distribution employs color difference analysis technology, statistically analyzing the color temperature values ​​of each region of the image in a standard color space. The dispersion of color temperature values ​​in different regions is calculated and compared with the target color temperature range. The detection of transition smoothness between adjacent frames is achieved by analyzing the trajectory changes of the sequence of images in the color space, quantifying the gradation characteristics of color parameters between frames. When color consistency deviates from the expected range, the system initiates a white balance reference point recalibration process, preferentially selecting a neutral gray region in the image as the new reference benchmark.

[0066] The backtracking optimization of noise reduction strategy parameters employs a hierarchical adjustment strategy. First, it checks the rationality of the chosen base filter type; then, it evaluates the applicability of the filter size and shape parameters; finally, it adjusts the execution order and intensity ratio of each filtering stage. The optimization process for spatial domain filtering rules considers the response characteristics of different noise features, establishing separate parameter adjustment rules for different types such as Gaussian noise, impulse noise, and stripe noise. The adjustment of temporal domain filtering parameters focuses on adapting the motion compensation intensity, dynamically optimizing the recursive filter weights based on the actual motion blur degree of the image sequence.

[0067] The white balance reference point calibration employs an intelligent selection algorithm that automatically identifies the most suitable neutral color region as a reference by analyzing the color distribution characteristics of the image content. The calibration process considers the semantic information of the scene, prioritizing object surfaces with stable color characteristics as references, such as concrete walls and asphalt roads. The color temperature range setting is automatically adjusted based on ambient lighting conditions, with a slightly wider tolerance range for cloudy scenes and a more stringent control standard under standard lighting conditions.

[0068] The detail retention assessment incorporates a local adaptive detection mechanism, assigning differentiated scoring weights to regions of varying importance within the image. Visually salient regions, such as building outlines and road signs, are given higher detail retention requirements, while homogeneous regions, such as the sky and water surfaces, are allowed relatively lenient standards. This zonal evaluation method effectively avoids the flaw of overall scoring masking local quality issues, ensuring that image details in critical areas are adequately protected.

[0069] The monitoring of color transition smoothness employs a joint temporal analysis method, examining not only the color gradation characteristics within a single frame but also tracking and analyzing the color evolution of the sequence of images over time. When a significant color jump is detected between adjacent frames, the system automatically checks the parameter settings of the color enhancement constraint boundaries, particularly the smoothness of the saturation adjustment curve and the continuity of the tone mapping. During optimization, the reversibility of the color enhancement process is maintained, ensuring that the adjusted image can recover its original color relationships through inverse transformation.

[0070] The dynamic adjustment of the processing flow establishes a priority mechanism, determining the optimization order based on the severity of quality issues. When multiple quality indicators are detected simultaneously, the system processes them sequentially in the order of signal-to-noise ratio defects, color deviation, and detail loss. After each optimization iteration cycle, the entire quality assessment process is re-executed, forming a closed-loop feedback adjustment. This incremental optimization method can achieve optimal quality improvement within limited computing resources.

[0071] The historical data recording and analysis function during parameter adjustment provides the system with learning capabilities. Each successful parameter optimization case is recorded in the knowledge base, forming processing experience for specific scenario types. When encountering a similar scenario again, the system prioritizes using the historically optimal parameter combination as the initial setting, significantly improving processing efficiency. The knowledge base adopts a scenario-feature-based classification storage structure, supporting fast retrieval and similarity matching.

[0072] The anomaly handling mechanism incorporates multi-level fault tolerance strategies. When automatic optimization fails to meet quality requirements even after reaching the maximum number of iterations, the system prompts for manual intervention and marks the case as a special sample for subsequent analysis. For specific scenario types where quality issues recur, the system automatically creates dedicated handling solutions to avoid impacting the normal processing flow of other scenarios. This intelligent anomaly management mechanism significantly improves the system's robustness and usability.

[0073] The final quality verification report contains comprehensive evaluation data, recording not only the quantitative results of various quality indicators but also highlighting key areas of concern in the images and providing processing recommendations. The report is presented visually, using heatmaps and other methods to intuitively display the distribution of quality issues, providing a clear reference for subsequent manual adjustments. The system supports saving successful combinations of processing parameters as preset solutions for direct use in similar scenarios, creating a virtuous cycle of continuous optimization.

[0074] This implementation achieves quality control over the UAV image processing workflow by establishing scientific evaluation standards and a flexible adjustment mechanism. The multi-layered design, from adjusting basic parameters to handling abnormal situations, ensures stable system performance in various complex scenarios. The introduction of continuous learning capabilities allows the system to continuously accumulate experience, gradually improve its automation level, and ultimately form a highly efficient and reliable image processing solution.

[0075] 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.

[0076] 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 denoising and enhancing color in UAV images, characterized in that, include: Acquire the original image sequence collected during the flight of the UAV, and identify the ambient light intensity change data and motion blur feature distribution corresponding to the original image sequence; Based on the ambient light intensity change data, the temporal dimension features of the original image sequence are divided, and combined with the motion blur feature distribution, a noise feature distribution map of the original image sequence is constructed. The device imaging parameter set of the original image sequence is invoked, the imaging mode constraints corresponding to the device imaging parameter set are parsed, and a multi-level noise reduction strategy framework for the original image sequence is formulated based on the imaging mode constraints and the noise feature distribution map.

2. The method for denoising and enhancing color of UAV images according to claim 1, characterized in that, The multi-level noise reduction strategy framework for the original image sequence, based on the imaging mode constraints and the noise feature distribution map, includes: Extract the high-frequency and low-frequency noise components from the noise feature distribution map, and calculate the energy distribution ratio of the high-frequency and low-frequency noise components in the time dimension; correlate the exposure time threshold and the photosensitive element sensitivity parameter in the imaging mode constraints to determine the suppression intensity level corresponding to the high-frequency noise component and the smoothing depth corresponding to the low-frequency noise component; based on the suppression intensity level and the smoothing depth, generate the spatial domain filtering rules and temporal domain filtering rules in the multi-level noise reduction strategy framework.

3. The method for denoising and enhancing color of UAV images according to claim 1, characterized in that, After constructing the noise feature distribution map of the original image sequence, the method further includes: The color space conversion data of the original image sequence is collected, and the color gamut coverage and color distortion probability distribution corresponding to the color space conversion data are analyzed. Based on the color gamut coverage, a reference color mapping relationship of the original image sequence is established. Combined with the color distortion probability distribution, the correctable parameter range of the reference color mapping relationship is calculated. Based on the correctable parameter range, a color enhancement constraint boundary for the original image sequence is generated.

4. The method for denoising and enhancing color of UAV images according to claim 3, characterized in that, After generating the color enhancement constraint boundary for the original image sequence, the method further includes: The multi-level noise reduction strategy framework is executed to process the original image sequence to obtain a set of intermediate images after noise reduction. The local contrast features and global average brightness of the intermediate image set are extracted and associated with the upper limit threshold of saturation and the hue balance coefficient in the color enhancement constraint boundary. Based on the local contrast features and the global average brightness, the applicable area range of the upper limit threshold of saturation in the intermediate image set is dynamically adjusted, and the weight distribution ratio of the hue balance coefficient in the shadow area and the highlight area is simultaneously corrected.

5. The method for denoising and enhancing color of UAV images according to claim 4, characterized in that, The dynamic adjustment of the applicable region range of the upper limit of saturation in the intermediate image set includes: The intermediate image set is divided into landform type regions and artificial building regions. The natural color distribution probability of the landform type region and the texture complexity index of the artificial building region are calculated. The vegetation cover sensitivity coefficient and water reflection feature parameters in the applicable area are correlated to establish a color correction priority sequence between the natural color distribution probability and the texture complexity index. Based on the color correction priority sequence, the first gradient adjustment step size of the saturation upper limit threshold in the landform type region and the second gradient adjustment step size in the artificial building region are configured.

6. The method for denoising and enhancing color in UAV images according to claim 1, characterized in that, After formulating the multi-level noise reduction strategy framework for the original image sequence, it also includes: The system calls upon a reference image database from historical flight missions and matches the scene similarity probability distribution between the reference image database and the current original image sequence. It then analyzes the matching degree of illumination conditions and the overlap degree of ground feature contours in the scene similarity probability distribution to generate a noise suppression reference template based on the reference image database. Finally, it fuses the noise suppression reference template with the spatial filtering parameters in the multi-level noise reduction strategy framework to optimize the motion blur compensation intensity in the temporal filtering rules.

7. The method for denoising and enhancing color of UAV images according to claim 4, characterized in that, The synchronous correction of the weight distribution ratio of the tone balance coefficient in the shadow and highlight areas includes: The boundary coordinates of the shadow region and the proportion of overexposed pixels in the highlight region are detected in the intermediate image set. The geometric continuity index of the boundary coordinates of the shadow region and the halo diffusion range of the overexposed pixel ratio are calculated. The shadow detail preservation factor and the highlight suppression attenuation parameter in the weight allocation ratio are associated to construct a tone correction linkage mechanism between the geometric continuity index and the halo diffusion range. Based on the tone correction linkage mechanism, the optimal tone compensation curve of the shadow region and the optimal brightness attenuation curve of the highlight region are output.

8. The method for denoising and enhancing color of UAV images according to claim 5, characterized in that, After configuring the upper limit threshold for saturation in the first gradient adjustment step size of the landform type region, the method further includes: Identify the vegetation cluster distribution characteristics and soil bare area range in the landform region, measure the chlorophyll reflectance intensity of the vegetation cluster distribution characteristics and the infrared absorptivity of the soil bare area range; correlate the vegetation health correction coefficient and soil moisture influence factor in the first gradient adjustment step size, and establish a saturation adaptive adjustment model between the chlorophyll reflectance intensity and the infrared absorptivity; dynamically control the execution frequency of the first gradient adjustment step size in dense vegetation areas and soil bare areas through the saturation adaptive adjustment model.

9. The method for denoising and enhancing color of UAV images according to claim 1, characterized in that, After obtaining the set of intermediate images after noise reduction, the process further includes: Verify the signal-to-noise ratio improvement and detail retention score of the intermediate image set. When the signal-to-noise ratio improvement is lower than a preset threshold, backtrack the spatial domain filtering rule parameters in the multi-level noise reduction strategy framework. Simultaneously detect the color consistency distribution and the smoothness of the transition between adjacent frames of the intermediate image set. When the color consistency distribution deviates from the target color temperature range, recalibrate the white balance reference point in the color enhancement constraint boundary.

10. A UAV image denoising and color enhancement optimization system, used to implement the UAV image denoising and color enhancement optimization method as described in any one of claims 1-9, characterized in that, The system includes: The image feature analysis module is used to acquire the original image sequence collected during the flight of the UAV, identify the ambient light intensity change data and motion blur feature distribution corresponding to the original image sequence, divide the time dimension features based on the ambient light intensity change data, and construct a noise feature distribution map by combining the motion blur feature distribution. The noise reduction strategy generation module is used to call the device imaging parameter set and parse the imaging mode constraints, and formulate a multi-level noise reduction strategy framework based on the imaging mode constraints and the noise feature distribution map. The color constraint construction module is used to collect color space conversion data, analyze the color gamut coverage and color distortion probability distribution, establish a baseline color mapping relationship and calculate the correctable parameter range, and generate color enhancement constraint boundaries. The collaborative optimization execution module is used to execute a multi-level noise reduction strategy framework to generate an intermediate image set, extract local contrast features and global brightness mean, dynamically adjust the applicable area range of the upper limit threshold of saturation and correct the weight allocation ratio of the tone balance coefficient. The output verification module is used to verify the signal-to-noise ratio improvement and detail retention score, backtrack the spatial domain filtering rule parameters, detect the color consistency distribution and the smoothness of the transition between adjacent frames, and calibrate the white balance reference point.

Images