A method and system for unmanned aerial vehicle multiscale image denoising

Abstract

The unmanned aerial vehicle multi-scale image denoising method and system provided by the application relates to the technical field of image processing. The multi-scale image denoising method and system are constructed, the problem that a real point light source and background noise are difficult to distinguish in night unmanned aerial vehicle inspection is effectively solved, and specifically, a multi-dimensional feature discrimination mechanism based on optical morphology, energy distribution and spatial context is introduced, and a strategy of dynamically calculating a boundary expansion coefficient in combination with a hierarchical depth difference is combined, so that fine protection of the real point light source is realized. Compared with the existing algorithm, the application can maintain the scale invariance of the signal in the multi-scale decomposition process, ensure that the depth denoising is performed at the same time, significantly improve the reconstruction sharpness and spatial feature integrity of the target signal (such as light and landmarks), and effectively reduce the multi-scale feature misjudgment rate.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and more specifically, to a method and system for multi-scale image denoising of unmanned aerial vehicles (UAVs). Background Technology

[0002] In urban nighttime drone patrol missions, due to insufficient ambient lighting, the energy distribution of real point light sources such as landmark traffic lights and streetlights in images often exhibits severe aliasing between the real light sources and background noise in scale space. Current mainstream multi-scale denoising algorithms often lack a scale-invariant signal discrimination mechanism for real point light sources during the layer extraction stage. This leads to real point light sources being incorrectly identified as persistent patch noise and smoothed, severely impacting the sharpness of the reconstructed image and the preservation of traffic light features. The core reason for this problem is that traditional algorithms only rely on a single dimension of amplitude or energy to remove noise, failing to fully utilize the optical morphology, energy attenuation patterns, and spatial contextual relationships of point light sources at different scales. Consequently, the algorithms cannot remove background noise while preserving important landmark signal features. Therefore, existing technologies urgently need improvement to address these issues. Summary of the Invention

[0003] This application discloses a multi-scale image denoising method and system for unmanned aerial vehicles (UAVs), aiming to solve the technical problem that features of real point light sources and background noise are easily aliased and incorrectly smoothed during multi-scale denoising in low-light environments at night. The technical solution of this application is as follows: In a first aspect, this application discloses a multi-scale image denoising method for unmanned aerial vehicles (UAVs), comprising: performing luminance channel decomposition on low-light nighttime images acquired by the UAV through non-downsampling transformation to construct a scale space containing multiple scale layers; identifying cross-scale stable high-response spots in the scale space, and identifying real point sources to be protected based on multi-dimensional features of optical morphology, energy, and spatial context; for the identified real point sources, delineating a core strong protection region and a surrounding weak protection region based on the centroid of the real point source, and generating an initial protection marker matrix for a seed layer; using the seed layer as the first scale layer, decomposing the initial protection marker matrix into a structured instruction package and transmitting it to the second scale layer, wherein the instruction contains the initial boundary parameters of the core strong protection region; calculating the boundary expansion coefficient of the second scale layer based on its level difference with the first scale layer; adjusting the initial boundary parameters using the boundary expansion coefficient to obtain target boundary parameters adapted to the second scale layer; and regenerating a dedicated protection marker matrix in the second scale layer based on the target boundary parameters to guide the denoising operation of the second scale layer. This technical solution enables multi-scale spatial localization and protection of point light source features, effectively distinguishing between real point light sources and persistent patch noise, and improving the quality of nighttime image reconstruction.

[0004] Secondly, this application also discloses a multi-scale image denoising system for unmanned aerial vehicles (UAVs), comprising: a luminance channel decomposition module, used to perform luminance channel decomposition on low-light nighttime images acquired by the UAV through non-subsampling transformation, constructing a scale space containing multiple scale layers; a real point light source identification module, used to identify cross-scale stable high-response spots in the scale space, and to identify real point light sources to be protected based on multi-dimensional features of energy, optical morphology, and spatial context; and an initial protection marker matrix generation module, used to delineate a core strong protection area and a surrounding weak protection area for the identified real point light sources, with the centroid of the real point light source as the reference. The system protects the region and generates an initial protection marker matrix for the seed layer. An inter-layer transfer and verification module is used to treat the seed layer as the first scale layer, decompose the initial protection marker matrix into structured instruction packages, and transfer them to the second scale layer. These instructions contain the initial boundary parameters of the core strongly protected region. The second scale layer calculates the boundary expansion coefficient based on its level difference with the first scale layer. The initial boundary parameters are adjusted using the boundary expansion coefficient to obtain target boundary parameters adapted to the second scale layer. Based on the target boundary parameters, a dedicated protection marker matrix is ​​regenerated in the second scale layer to guide the denoising operation.

[0005] Beneficial Effects: This application effectively solves the problem of distinguishing between real point light sources and background noise in nighttime UAV inspections by constructing a multi-scale image denoising method and system. The scheme introduces a multi-dimensional feature discrimination mechanism based on optical morphology, energy distribution, and spatial context, and combines it with a strategy of dynamically calculating the boundary expansion coefficient based on hierarchical depth differences, achieving refined protection of real point light sources. Compared to existing algorithms, this application can maintain the scale invariance of the signal during multi-scale decomposition, ensuring that while achieving depth denoising, it significantly improves the reconstruction sharpness and spatial feature integrity of target signals (such as lights and landmarks), effectively reducing the multi-scale feature misjudgment rate. Attached Figure Description

[0006] To illustrate this application more clearly, the accompanying drawings used in the embodiments will be briefly described below. Obviously, those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0007] Figure 1 The diagram above illustrates the overall workflow architecture of a multi-scale image denoising method for unmanned aerial vehicles (UAVs). Figure 2 The diagram above illustrates a flowchart of a multi-scale image denoising method for unmanned aerial vehicles (UAVs). Figure 3 An exemplary schematic diagram of a multi-scale image denoising system for unmanned aerial vehicles is shown.

[0008] Reference numerals: 100, A multi-scale image denoising system for UAVs; 10, Luminance channel decomposition module; 20, Real point light source identification module; 30, Initial protection mark matrix generation module; 40, Interlayer transfer and verification module. Detailed Implementation

[0009] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0010] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0011] In urban nighttime drone patrol missions, drones often need to acquire images in low-light environments. Due to the dim ambient light, image sensors inevitably introduce a large amount of background noise. To improve image quality, multi-scale denoising algorithms are typically used to process the images. However, urban nighttime scenes contain a large number of regularly distributed point light sources such as landmark traffic lights and streetlights. The cross-scale response characteristics of these point light sources under multi-scale decomposition are highly similar to persistent patch noise. Existing multi-scale denoising algorithms lack a discrimination mechanism for scale-invariant signals in the layered extraction and processing stages, often misclassifying these real point light sources as noise and smoothing them, resulting in blurred edges, energy attenuation, or even complete loss of traffic lights or streetlights in the image. This phenomenon of energy distribution aliasing between landmark traffic lights and background noise in the scale space directly exposes the core defects of existing technologies in feature protection and discrimination mechanisms.

[0012] To more clearly explain the technical solution of this application, the basic concepts involved are explained here. "Scale space" refers to a set of images decomposed into different frequency bands through specific mathematical transformations of the original image. Each frequency band corresponds to a scale layer, reflecting the features of the image at different levels of detail. More specifically, the original nighttime image contains both fine and rapidly changing high-frequency information, such as the edges of streetlights, and slowly changing low-frequency information, such as road outlines, building shapes, and large-scale halos. After decomposing the image into scale space, different types of information are assigned to different scale layers, facilitating separate analysis. For example, the edge of a streetlight's wick is usually more concentrated in higher-frequency scale layers, while its outward-spreading soft halo is more likely to appear in lower-frequency scale layers. "Cross-scale stable high-response spots" refer to local bright spot regions that exhibit high brightness or energy response in multiple consecutive scale layers and whose positions are relatively fixed. The phrase "relatively fixed position" here does not require that the pixel coordinates in each scale layer be completely consistent. Instead, it allows for small-scale offsets caused by decomposition filtering, slight jitter, or imaging blur. As long as the offset is within the preset tolerance range, it can be considered a continuous response of the same candidate spot in different scale layers. For example, the same street lamp may exhibit multiple response regions with basically overlapping center positions and slightly varying bright spot ranges in adjacent scale layers. Such targets are usually those that need to be carefully identified. Random noise, on the other hand, often appears only occasionally in individual scale layers or has significant positional jumps and does not possess this continuous stability. The seed layer refers to a specific scale layer selected as a reference among many scale layers. It is used to initially delineate the light source area that needs protection and to transmit protection instructions to other scale layers based on this. The seed layer is usually selected from scale layers with relatively complete point light source structure, clear edge and halo transition relationships, and where noise interference has not been excessively amplified. In other words, the seed layer is not arbitrarily designated but is preferentially selected as the layer that best describes the spatial morphology of the real point light source stably. Taking urban nighttime street light scenarios as an example, if the bright spot in the center of the street light is clear, the outer diffusion area is identifiable, and the number of isolated noise points is relatively small in a certain scale layer, then that scale layer is more suitable as a seed layer.

[0013] like Figures 1-2 As shown, the technical solution disclosed in this embodiment is as follows: A method for multi-scale image denoising of unmanned aerial vehicles (UAVs) includes: S1, the brightness channel of the low-light nighttime image acquired by the UAV is decomposed by non-subsampling transformation to construct a scale space containing multiple scale layers; S2 identifies cross-scale stable high-response spots in scale space and identifies real point sources to be protected based on multi-dimensional features of optical morphology, energy, and spatial context. S3. For the identified real point light sources, the core strong protection area and the surrounding weak protection area are delineated based on the centroid of the real point light source, and the initial protection mark matrix of the seed layer is generated. S4, the seed layer is used as the first scale layer, the initial protection mark matrix is ​​decomposed into a structured instruction package and passed to the second scale layer, where the instruction contains the initial boundary parameters of the core strong protection region; S5, the second-scale layer calculates the boundary expansion coefficient based on the layer difference between it and the first-scale layer; the initial boundary parameters are adjusted using the boundary expansion coefficient to obtain the target boundary parameters adapted to the second-scale layer; and based on the target boundary parameters, a dedicated protection marker matrix is ​​regenerated in the second-scale layer to guide the denoising operation of the second-scale layer.

[0014] The overall operational logic of the above technical solution is as follows: First, the low-light nighttime image is decomposed into scale layers of different frequencies. Within these scale layers, it is searched and confirmed which bright spots are genuine point light sources (such as streetlights) rather than noise. To obtain the image to be processed, an inspection drone equipped with a visible light camera can continuously acquire image frames while flying along a preset route at night. These image frames can come from single-frame static images stored on an onboard storage device or from keyframes extracted at time intervals from a video stream. If the drone simultaneously records attitude and positioning information, this additional information can also be used for subsequent spatial context analysis, such as assisting in determining whether multiple candidate bright spots are continuously distributed along the road direction. After confirmation, in a baseline seed layer (i.e., the first scale layer), core areas requiring key protection and peripheral areas requiring transitional protection are delineated for these light sources, forming initial protection markers. Subsequently, this protection information is packaged and passed to other scale layers (i.e., the second scale layer). Since different scale layers represent different frequency details, and the diffusion range of the light source varies across different scale layers, the second scale layer calculates an expansion coefficient based on its own level difference with the seed layer. This coefficient adaptively adjusts the protection boundary, ultimately generating a protection marker specifically for that layer. This allows for precise avoidance or reduction of smoothing processing on the light source region during denoising. Furthermore, during denoising, the protection marker matrix can be used as the basis for adjusting the filtering intensity of that scale layer: locations within the core strong protection region retain their original response or undergo only very weak smoothing; locations within the peripheral weak protection region undergo restricted smoothing; and locations within the non-protected region undergo regular denoising. After this processing, the center brightness, edge sharpness, and peripheral halo transition of the real point light source can be preserved relatively completely in the final multi-scale reconstructed image.

[0015] To understand the core technical methods in the above solution, please refer to the following example: When decomposing the luminance channel using non-subsampling transform, to avoid image color distortion caused by the denoising process, the RGB format image captured by the drone can be pre-converted to the HSV color space, and only the V channel (luminance channel) can be extracted for processing. This is because the main noise interference in nighttime inspection images is usually manifested primarily in luminance fluctuations, while color information is more used for subsequent display and recognition. Therefore, prioritizing multi-scale denoising in the luminance channel can suppress noise while minimizing damage to the original color relationships. In one possible example, non-subsampling contourlet transform is used to decompose the luminance channel into a multi-level decomposition, resulting in a low-frequency approximation layer and multiple high-frequency detail layers, which together constitute the scale space. Here, "non-subsampling" means that the image size is not reduced during the decomposition process, so each scale layer maintains the same spatial size as the original image. This way, when transferring the protected area between different scale layers, no additional complex coordinate scaling calculations are needed; centroid positioning, boundary expansion, and marker matrix generation can be directly completed around the same image coordinate system. To make the decomposition results more stable, basic preprocessing can be performed on the input image, such as bad pixel correction, overexposed pixel suppression, or mild brightness normalization, to reduce the interference of extreme abnormal pixels on subsequent cross-scale response analysis.

[0016] When identifying and distinguishing real point light sources, the algorithm searches for local extrema in various high-frequency detail layers, extracting those spots whose coordinate positions deviate slightly across multiple adjacent scale layers and whose response values ​​are consistently higher than the background. This "consistently higher than the background" is achieved through local neighborhood comparison. A neighborhood window is set around a candidate spot, and the difference between the center response and the neighborhood average response, neighborhood median response, or neighborhood background estimate is compared. Only when this difference meets a preset judgment condition is the spot retained as a candidate. More specifically, in the context of urban streetlights, the spatial context features of these spots can be extracted, such as detecting whether these spots exhibit a collinear arrangement along both sides of the road in the world coordinate system; simultaneously, optical morphological features can be extracted, such as calculating the fit between the spot brightness distribution and the standard two-dimensional Gaussian point spread function.

[0017] To obtain arrangement information in the world coordinate system, pose data, camera intrinsics, and geolocation information recorded during drone photography can be used to project candidate spots in the image onto a ground reference coordinate system. Then, it can be analyzed whether multiple candidate spots are approximately arranged along the same road direction, or whether they are distributed at relatively uniform intervals on one or both sides of the road. If some bright spots are very bright in the image but are spatially isolated and disordered, and have no obvious connection with the road structure, they are more likely to be reflective noise, glass reflection, or random interference. For the extraction of optical morphological features, local image patches can be cropped around the candidate spots to analyze whether the distribution of a bright center with gradually fading edges is continuous and symmetrical, and whether there are obvious tails, breaks, or irregular blocky diffusions.

[0018] Real streetlights typically possess a relatively stable center-periphery attenuation structure, while patch noise is more prone to exhibiting characteristics such as fragmented boundaries, irregular shapes, and poor cross-scale consistency. By comprehensively verifying these multi-dimensional features, random patch noise can be accurately eliminated, preserving genuine streetlight sources. Furthermore, this comprehensive verification of multi-dimensional features can be performed using a step-by-step screening approach: first, cross-scale stability is used to eliminate most occasional noise; then, optical morphology is used to filter out abnormal bright spots that do not conform to point diffusion characteristics; and finally, spatial context is used to retain candidate targets consistent with the layout of roads and landmarks. This selection order, from easy to difficult, reduces the computational burden of subsequent complex analyses while improving the stability of identifying genuine point light sources.

[0019] When delineating the protection area and generating the initial protection marker matrix, the centroid coordinates of the identified streetlights are used as the center. In one possible example, a smaller radius is set as the core strong protection area, and a larger concentric annular area is set as the surrounding weak protection area. The centroid coordinates here can be obtained by weighting the center of high-response pixels within the candidate bright spot area, or by jointly determining the local brightness peak position and the neighborhood distribution. If a streetlight exhibits slight elliptical diffusion in the seed layer, rather than an ideal circular bright spot, the center of its main bright area can still be used as the centroid, and the protection boundary can be determined based on its lateral and longitudinal diffusion range.

[0020] For ease of implementation, the generated initial protection marker matrix can use a two-dimensional matrix with the same size as the seed layer image to record the protection status pixel by pixel. Each position in the matrix corresponds to a pixel position in the seed layer. If the position falls into the core strong protection area, a strong protection marker is written; if it falls into the surrounding weak protection area, a weak protection marker is written; if it does not belong to any protection area, a background marker is written. In the generated initial protection marker matrix, pixel positions within the core strong protection area are assigned a protection weight value of 1.0 (representing no denoising or smoothing), pixel positions within the surrounding weak protection area are assigned a protection weight value of 0.5 (representing mild denoising and smoothing), and the background area is assigned 0 (representing regular denoising). When the protection areas of multiple real point light sources overlap, they can be merged according to the principle of protection strength priority, that is, overlapping positions are given priority to retain higher-level protection markers, thereby avoiding protection weakening caused by mutual overlap. This method of layered marking between the core area and the surrounding area can clearly distinguish between two different processing requirements: the wick must be preserved and the halo can be moderately denoised, providing a direct basis for differentiated denoising at subsequent scales.

[0021] When calculating the boundary expansion coefficient at the second-scale layer, it needs to be determined based on the layer difference between the current layer and the first-scale layer (seed layer). In a possible example, a discrete mapping table of layer difference and expansion coefficient can be pre-constructed. This mapping table can be obtained through offline calibration, i.e., first selecting several representative nighttime patrol image samples, including streetlights or traffic lights under different brightness levels, shooting heights, and road types; then observing the diffusion range changes of the same real point light source in each scale layer, recording the boundary magnification required to fully cover the light source core and surrounding halo under different layer differences; finally, organizing the layer difference and the corresponding boundary expansion coefficient into a mapping relationship and storing it. For example, it can be recorded in the form of "when the layer difference is A, the corresponding expansion coefficient is B". After receiving the structured instruction package containing the initial boundary parameters, the second-scale layer directly calls the discrete mapping table to obtain the corresponding expansion coefficient, thereby completing the boundary adjustment. In addition to the boundary parameters of the initial core protected area, the structured instruction package can also contain information such as centroid coordinates, boundary parameters of the surrounding weakly protected areas, protection level identification, and target light source number. The advantage of using structured encapsulation is that different scale layers can directly parse the required fields when receiving protection information, without having to infer the protection area from the seed layer image again, which is more conducive to the consistency and real-time performance of cross-layer transmission.

[0022] Based on the above working principle, this overall technical solution achieves the following technical effects: By introducing a real point light source discrimination mechanism based on multi-dimensional features in the scale space and innovatively adopting a cross-layer transmission of protection instructions, the denoising algorithm can accurately distinguish between regular landmark light sources and persistent patch noise in urban nighttime scenes. Simultaneously, by dynamically adjusting the protection boundaries at different scale layers through boundary expansion coefficient calculation, the protected area can precisely match the physical diffusion morphology of the light source in different frequency bands. The underlying reason is that although real point light sources exhibit different spatial diffusion ranges in different frequency bands after multi-scale decomposition, their central location, overall attenuation trend, and structural relationship with the surrounding scene usually have strong continuity; while random noise lacks this continuous and stable cross-scale structure. The processing chain of discrimination, protection, and cross-layer expansion effectively marks areas that cannot be treated as noise before denoising, thus preventing subsequent filtering from indiscriminately covering these key areas. This closed-loop logic ensures that while strongly suppressing background noise in nighttime images, it fully preserves the high contrast and sharp edges of point light sources such as streetlights, and maintains a natural transition in the outer halo, thus solving the technical problem of light source blurring caused by traditional multi-scale denoising algorithms.

[0023] In some preferred embodiments, the step of calculating the boundary spread coefficient of the second-scale layer based on its hierarchical difference with the first-scale layer includes: The second-scale layer calculates the boundary expansion coefficient using a monotonically increasing function based on the difference in layer depth between itself and the first-scale layer. The difference in layer depth is the difference between the layer index of the second-scale layer and the layer index of the first-scale layer. The boundary expansion coefficient increases with the increase of the layer depth difference, causing the boundary range of the protected area to monotonically expand as the frequency band decreases.

[0024] In the aforementioned application scenarios, if a pre-constructed discrete mapping table is used to obtain the expansion coefficients, it is necessary to manually calibrate and input specific coefficient values ​​for each possible layer difference. This approach is not only cumbersome to configure when dealing with complex images with many decomposition layers and deep scale spaces, but the discrete coefficient values ​​also fail to smoothly reflect the gradual diffusion pattern of the light source halo across continuous frequency bands. To make the boundary expansion process more adaptive and conform to the laws of optical physics, this specific implementation scheme is proposed. The "monotonically increasing function" mentioned here can be understood as a conversion rule from continuous layer differences to expansion coefficients. Its basic requirements are: the closer the second scale layer is to the lower frequency end relative to the first scale layer, the larger the boundary expansion coefficient; when the layer difference is small, the expansion amplitude is relatively mild; when the layer difference continues to increase, the expansion amplitude continues to increase, but without sudden jumps. The purpose of adopting this continuous rule is not to pursue complex calculations, but to ensure that the protective boundary exhibits a smooth and predictable trend during cross-layer propagation.

[0025] The core of this specific implementation scheme lies in the introduction of a monotonically increasing function based on the hierarchical depth difference. More specifically, the hierarchical index represents the depth position of each frequency band in the scale space, with higher frequency layer index values ​​and lower frequency layer index values. The hierarchical index of the first scale layer (seed layer) can be denoted as L1, and the hierarchical index of the second scale layer can be denoted as L2. The hierarchical depth difference is obtained by subtracting the hierarchical index of the first scale layer from the hierarchical index of the second scale layer. In a possible example, an exponential monotonically increasing rule can be used to calculate the boundary spread coefficient, where the diffusion control constant is a preset parameter. The preset parameters can be obtained through sample calibration: first, select a batch of nighttime drone inspection images containing real point light sources with different brightness, shooting distances, and background complexities; then, determine the reasonable protection boundaries for these point light sources at each scale layer using manual verification or semi-automatic annotation; subsequently, compare the expansion trends of the protection boundaries under adjacent layer differences, and select the parameter range that allows for smooth boundary changes without excessively encroaching on the background area; when the light source core remains intact, the halo transition is natural, and background noise is effectively suppressed in multiple sets of samples, the corresponding parameters can be solidified as diffusion control constants. The reason for using an exponential monotonically increasing rule is that the diffusion of nighttime point light sources in lower frequency scale layers is often not linearly increasing, but rather closer to a trend of slower changes in the early stages and more obvious diffusion in the later stages. Therefore, this type of rule is more likely to fit the actual halo diffusion characteristics. Of course, without deviating from the core principle of increasing with the layer depth difference, other continuously increasing rules can also be used, as long as the smooth and stable expansion of cross-layer boundaries can be guaranteed.

[0026] Through the above technical solution, the algorithm utilizes a monotonically increasing function to transform discrete hierarchical index differences into continuous and smooth expansion coefficients. Since the physical distribution range of low-frequency components of point light sources in multi-scale image decomposition is typically larger than that of high-frequency components, the characteristic of the coefficients calculated by this function monotonically increasing with decreasing frequency band aligns with this optical diffusion law. Further analysis of its mechanism reveals that if the expansion coefficient changes too abruptly, obvious discontinuities will appear at the protection boundaries between adjacent scale layers, ultimately leading to unnatural brightness transitions during image reconstruction. The continuously increasing expansion method allows the protection area in each scale layer to expand outwards layer by layer, forming a protection structure that gradually connects from the center to the periphery. The final derived technical effect is that the algorithm can automatically and smoothly calculate the boundary expansion coefficients required for any scale layer without needing to separately look up tables for each hierarchical difference. This significantly improves the adaptability and robustness of the cross-scale protection mechanism in different depth scale spaces, while reducing parameter maintenance costs.

[0027] In practical applications, when adjusting the initial boundary parameters using the boundary expansion coefficient, the initial core protection radius and the initial peripheral protection radius are multiplied by the boundary expansion coefficient to obtain the target core protection radius and the target peripheral protection radius. The initial core protection radius corresponds to the scale of the core strong protection area in the first scale layer, and the initial peripheral protection radius corresponds to the scale of the peripheral weak protection area in the first scale layer. The values ​​of the target core protection radius and the target peripheral protection radius both increase with the increase of the layer depth difference.

[0028] After obtaining the boundary expansion coefficient, a crucial step is applying it to the initial boundary parameters. Simply increasing the value uniformly on the outermost boundary would disrupt the geometric proportions between the core strongly protected region and the surrounding weakly protected regions, potentially leading to abrupt step artifacts between the light source center and the halo transition zone during image reconstruction. To maintain the structural proportions within the protected region, this specific implementation scheme is proposed. The "multiplying by the boundary expansion coefficient separately" essentially performs simultaneous scaling on both the core strongly protected region and the surrounding weakly protected regions at the same scale, rather than scaling only one region or applying independent scaling rules to both. This ensures that the core region remains within the surrounding region, and the relative hierarchical relationship between them is not disrupted by scale changes.

[0029] This specific implementation plan clarifies the execution logic of boundary adjustment, namely, the use of an independent multiplicative scaling mechanism. For example, in the first scale layer, the initial core protection radius passed from the structured instruction packet is Rcore, and the initial peripheral protection radius is Rperi, where Rperi is greater than Rcore. When the second scale layer calculates the boundary expansion coefficient K, the algorithm performs proportional scaling on the initial core protection radius and the initial peripheral protection radius to obtain the target core protection radius and the target peripheral protection radius. If the calculation result shows non-integer pixel boundaries, rounding, nearest neighbor interpolation, or boundary rasterization can be used to convert them into the actual pixel range that can be written into the protection marker matrix. Furthermore, when generating the dedicated protection marker matrix for the second scale layer based on the target radius, the pixel positions in this layer of the image can still be re-traversed using the original center coordinates as the center, determining whether each pixel falls within the target core protection radius, falls within the target peripheral protection radius but is outside the core area, or is outside the protection area, and writing the corresponding protection weight accordingly. If a real point light source exhibits significant non-circular symmetric diffusion in a lower frequency scale layer, the circular boundary can be replaced with an elliptical boundary or a directional boundary, provided that the principle of "synchronous expansion of the core area and the surrounding area" remains unchanged, as long as the core area and the surrounding area maintain the same expansion logic.

[0030] By multiplying the radii of both the core and periphery dimensions by the same expansion coefficient, the structural ratio between the strong core protection area and the weak periphery protection area remains stable across all scale layers. This proportionally scaled working principle ensures that the protective mask for the light source maintains a similar geometric shape across different frequency bands. Further derivation shows that when the multi-scale denoising results are finally reconstructed back into the original image, the protective effects of each scale layer on the same light source will be superimposed around the same center and in similar shapes. Therefore, there will be no imbalance where one layer only protects the light wick while another excessively intrudes into the background, or one layer provides excessively wide halo protection while another provides excessively narrow protection. The resulting technical effect is that in the final reconstruction stage of nighttime drone images, the bright center of the light source can naturally and smoothly transition to the surrounding weak halo area, avoiding ringing effects or harsh artifact boundaries caused by disproportionate protection area ratios at the light source edges, significantly improving the visual naturalness of landmark light sources.

[0031] In urban nighttime scenes, streetlights or traffic lights are often densely arranged. When regenerating the dedicated protection marker matrix at the second scale layer (especially the low-frequency layer, after the protection boundary is extended), the weak protection regions around adjacent light sources are prone to overlap. If these overlapping regions are not specially processed, it may lead to uncontrolled denoising intensity in the overlapping regions, or protection conflicts, making it impossible to effectively remove background noise between dense light sources, or causing abnormal artifacts at the light source edges. To solve the denoising conflict problem caused by the overlap of protection boundaries of dense light sources, further optimization is made based on the boundary extension mechanism mentioned above. The weak protection region mentioned here refers to a transitional protection region set around the outer edge of the core strong protection region of the real point light source. This region is not completely prohibited from filtering, but allows for a weaker degree of smoothing under controlled conditions to balance halo preservation and noise removal. The "dedicated protection marker matrix" is a marker data structure that records the protection status of each pixel in the current scale layer, and is used to distinguish at least the unprotected region, the core strong protection region, the weak protection region, and the region where boundary conflicts occur. In nighttime drone-collected images, especially in scenarios such as continuous streetlights on main roads, multiple traffic lights at intersections, and light strings on commercial street signs, the halos of different light sources are more likely to spread outward and overlap with each other in the low-frequency scale layer. Therefore, boundary conflicts are not occasional occurrences, but rather high-frequency situations that require specially designed conflict handling rules.

[0032] In some preferred embodiments, the method further includes: when regenerating the dedicated protection marker matrix at the second scale layer, detecting the existence of boundary conflict states pixel by pixel; a boundary conflict state refers to a pixel coordinate falling simultaneously into the weak protection regions surrounding two or more different real point light sources, and the pixel falling into the overlapping region is a conflicting pixel; in specific implementation, based on the centroid coordinates of the real point light sources, the corresponding protection radius, and the expanded boundary range confirmed in the previous stage, an independent region occupancy record can be generated for each real point light source at the second scale layer, and then the pixels are traversed and checked according to the image scanning order. When a pixel is written into the dedicated protection marker matrix, it is detected that it has been simultaneously associated with the weak protection region identifiers surrounding two or more different light sources, that is, the pixel is marked as a conflicting pixel. The pixel-by-pixel detection here can be implemented by software, for example, when the processor executes the image processing program, accessing the image coordinates row by row and column by column; it can also be implemented by a parallel image processing unit, for example, dividing the image into multiple sub-blocks for parallel checking and then merging the conflict results.

[0033] In some embodiments, the second scale layer image used in this step can be obtained from night scene images continuously captured by a visible light camera mounted on a drone, after distortion correction, grayscale normalization, and multi-scale decomposition. If a boundary conflict exists, the mild smoothing protection threshold corresponding to each light source causing the conflict is extracted, and the controlled smoothing threshold of the conflicting pixel is set to the maximum value among the mild smoothing protection thresholds corresponding to all light sources. The value is used to perform controlled smoothing on the decomposition coefficients of the conflicting pixel. The mild smoothing protection threshold is a threshold used to control the noise suppression intensity within the surrounding weak protection area, and only coefficients exceeding this threshold are subjected to mild contraction towards zero. In specific implementation, when generating the surrounding weak protection area for each real point light source, a mild smoothing protection threshold field can be pre-written according to the local noise level near the light source, the halo brightness attenuation state, and the frequency band characteristics of the scale layer to which it belongs. When a conflicting pixel is identified, the system no longer executes the smoothing instructions for multiple light sources separately, but first reads all the mild smoothing protection thresholds associated with the pixel, and then selects the one with the largest value as the only effective controlled smoothing threshold for the pixel. This means that conflicting pixels are handled according to the most lenient protection standard. That is, only when the decomposition coefficient significantly exceeds the threshold will a slight reduction be performed; otherwise, it remains unchanged. This avoids excessive suppression caused by the superposition of multiple smoothing instructions. The "slight reduction towards zero" mentioned here can be understood as reducing the decomposition coefficient above the threshold by only a limited amount, rather than clearing it to zero, in order to preserve the effective structural information carried by the light source edge and halo gradient.

[0034] This specific implementation plan clarifies the detection mechanism for boundary conflict states and the execution logic of controlled smoothing operations. For example, consider the following scenario: Assume there are two adjacent LED streetlights, A and B, in an image of a main urban road captured by a drone. In the lower-frequency second-scale layer, due to the boundary expansion coefficient, the weakly protected areas surrounding streetlights A and B intersect, forming an overlapping region. When generating a dedicated protection marker matrix pixel by pixel, the algorithm detects that the pixel coordinates within this overlapping region are simultaneously associated with the protection commands of both streetlights A and B, thus triggering a boundary conflict state. At this point, the algorithm extracts the mild smoothing protection threshold (e.g., set to T_A) for streetlight A and the mild smoothing protection threshold (e.g., set to T_B) for streetlight B. By comparison, the maximum value of the two (assuming T_A is greater than T_B, then T_A is used) is selected as the final controlled smoothing threshold for the conflicting pixel. In subsequent denoising filtering, only when the decomposition coefficient magnitude of the pixel exceeds T_A is a mild shrinkage process towards zero performed; otherwise, the original value is retained. To further illustrate this process, the overlapping area can be understood as a transitional light-emitting band between two streetlights. In the original image, this area contains both the true brightness variation formed by the superposition of halos from both sides and random noise introduced by the sensor gain at night. If each streetlight (A) and streetlight (B) is processed independently according to its own threshold, the smaller threshold will take precedence, thus attenuating the transitional brightness that should be preserved as noise. By uniformly adopting a larger threshold, the system can treat this area as a "highly protected intersecting halo region," only suppressing noise components that significantly deviate from the normal brightness distribution. Therefore, it is more suitable for the true imaging characteristics of continuously arranged light sources in urban night scenes.

[0035] The working principle of this specific implementation scheme is to ensure that when multiple light source protection requirements conflict, the system automatically adopts the most lenient protection standard (i.e., allowing more high-frequency details to be preserved) by using a "maximum value" strategy. This processing logic effectively avoids excessive suppression of the halo of adjacent light sources due to an excessively small smoothing threshold for a certain light source. The resulting technical effect is that in scenarios with dense streetlights or complex landmark light sources, overlapping halo areas can achieve natural and smooth visual fusion, completely solving the problems of boundary fragmentation, artifact abrupt changes, and uncontrolled denoising intensity that are easily caused by traditional denoising algorithms when dealing with dense light sources. The reason can be further explained as follows: When multiple weakly protected areas overlap, it essentially means that the location simultaneously carries energy information diffused outward from multiple real light sources. If a strict threshold is still used for suppression, this portion of real diffused energy will be mistakenly deleted, ultimately resulting in brightness discontinuities or edge gaps in the reconstructed image. However, by using the maximum mild smoothing protection threshold, the priority of preserving conflicting areas is increased, and the brightness transition formed by the superposition of multiple sources can be completely preserved. Only noise components that significantly exceed the normal distribution are contracted. Therefore, the image appears as having natural halo connections, continuous edge transitions, and minimal residual background noise. Especially when using drones to capture aerial views of roads, bridges, or large squares, regularly arrayed lights can form repetitive and dense bright spots in the image. This solution can significantly improve the overall visual consistency of these complex light source groups.

[0036] In multi-scale denoising, the scale space contains multiple scale layers. If each scale layer independently calculates and matches the location and protection range of the real point light source, the computational load is enormous, and positioning errors are easily caused by different noise distributions in each layer. To achieve efficient and accurate transmission of protection information between different scale layers, avoid redundant calculations, and ensure consistency of cross-layer protection, a standardized information encapsulation and transmission mechanism is needed. Here, "scale space" refers to the set of multiple frequency band levels obtained after multi-scale decomposition of the same input image, with each layer corresponding to a different degree of detail preservation. "Consistency of cross-layer protection" means that the protection center location, protection boundary generation basis, and protection level instructions of the same real point light source in different scale layers maintain a unified source, avoiding one layer identifying it as a real light source while another layer is mistakenly offset or misses detection due to local noise differences. The input image can be obtained by continuously capturing static image frames from a UAV equipped with a gimbal camera during nighttime inspection flights, or by extracting frames from a nighttime video stream, aligning them with timestamps, and then sending them into the same multi-scale denoising pipeline.

[0037] Optionally, the structured instruction package includes: the hierarchical index of the first scale layer, the centroid coordinates of each confirmed real point light source in the first scale layer, the initial core strong protection region radius, and the initial peripheral weak protection region radius. As a lightweight metadata package, the structured instruction package is broadcast to all other target scale layers in the scale space via a cross-layer communication mechanism. The "structured instruction package" described here is not a complete image mask, but rather a set of parameters after compressing and expressing the protection information. Its recorded content is sufficient to allow other target scale layers to recover the protection marker matrix of their current layer based on the same reference without re-performing complete light source identification. The centroid coordinates can be represented using horizontal and vertical coordinates in the image coordinate system, while the initial core strong protection region radius and the initial peripheral weak protection region radius describe the original protection boundary size of the light source under the seed layer. In actual encapsulation, in addition to the above basic fields, a unique light source identifier, source frame number, confidence level, or timestamp can also be included for consistency verification during multi-frame processing and cross-module calls. However, these additional fields do not change the basic meaning of the structured instruction package, which focuses on the hierarchical index, centroid coordinates, and protection radius. The so-called "cross-layer communication mechanism" can be implemented by a shared memory area within the same process, or by passing message queues, cache objects, or function call parameters between modules, as long as it can ensure that other target scale layers receive the structured instruction packet before generating the protection tag matrix of this layer.

[0038] This specific implementation plan details the data composition of the structured instruction package and the operation of the cross-layer communication mechanism. For example, after completing the light source identification and boundary delineation at the seed layer, the algorithm encapsulates the relevant parameters into a lightweight meta-data package in JSON or a specific structure format. This data package explicitly records the layer index of the seed layer (e.g., L=2), the set of centroid coordinates of multiple real point light sources identified in the image, and the corresponding initial core strong protection region radius and initial peripheral weak protection region radius. After encapsulation, the structured instruction package is broadcast to all other target scale layers (layers 1, 3, 4, etc.) in the scale space using cross-layer communication mechanisms such as shared memory or message queues within the system. Upon receiving the instruction package, each target scale layer directly parses the baseline parameters and, combined with its own layer difference from layer 2, independently calculates the expansion coefficients and generates the protection marker matrix for its own layer. For ease of implementation, a specific encapsulation example can be given: For a given real point light source, the structure can sequentially record the "light source number," "centroid horizontal coordinates," "centroid vertical coordinates," "initial core strong protection zone radius," "initial surrounding weak protection zone radius," and "seed layer index." Once multiple light sources are identified, the system combines these records into an array or list and sends them as a unified package. During parsing, the receiving end does not need to re-access the original image for peak search; instead, it directly uses the centroid coordinates in the structured instruction packet as the center point, the corresponding protection radius as the reference value, and combines this with the extension rules of this scale layer to generate a dedicated protection marker matrix for this layer. If the system adopts a continuous video stream processing mode, the same structured instruction packet can also be bound to the current frame number to prevent the protection parameters of the previous frame from being mistakenly used in the next frame.

[0039] The working principle of this specific implementation scheme lies in abstracting and reducing the complex image mask matrix into a set of core parameters. This transforms the feature matching and boundary calculation process, which originally required repeated execution layer by layer, into a parameter parsing process that is calculated once and shared across multiple layers. The resulting technical effects are: significantly reducing the computational overhead and memory usage of multi-scale denoising algorithms when transmitting protection information across layers; and ensuring absolute spatial alignment of protection operations for the same light source across different frequency bands in the scale space based on a unified reference parameter broadcasting mechanism, avoiding positioning drift caused by differences in noise distribution across layers, and significantly improving the overall efficiency and stability of the denoising system. This is because directly transmitting the complete mask across layers not only involves a large amount of data, but also requires re-adaptation if there are slight changes in the image size, resolution mapping relationship, or boundary expansion state of different layers. However, by using structured instruction packets, only the most critical information such as "which layer is the reference, where the light source center is, and how large the basic boundary is" is retained. Other target scale layers can then automatically reconstruct the protection range according to unified rules. Therefore, this reduces the transmission burden and ensures that each layer expands from the same origin and the same initial scale. Furthermore, since the confirmation of real point light sources is only completed in the seed layer, other layers avoid repeatedly performing operations such as candidate spot extraction, confidence verification, and boundary fitting. Therefore, the overall processing chain is more stable, which is especially suitable for nighttime mission scenarios where drone platforms are more sensitive to processing latency and resource consumption.

[0040] In the aforementioned cross-layer transmission mechanism, the selection of the first scale layer (i.e., the seed layer) is crucial. If the selected seed layer frequency is too high, it may contain a large amount of unfiltered high-frequency noise, leading to misjudgment; if the selected seed layer frequency is too low, the edge of the light source has been severely diffused, making it impossible to accurately locate the centroid and initial boundary. To automatically find the most suitable scale layer as a benchmark in scale space and ensure the accuracy of the initial protection marker matrix, a quantitative selection criterion for the seed layer is proposed. The "high-frequency energy purity" mentioned here can be understood as the clarity of the real point light source response relative to the background noise within a certain scale layer. The higher the value, the more prominent the effective light source information in the candidate spot region and the less interference from background noise. "Identification stability" reflects whether there are enough real point light sources that can be stably identified in the scale layer. If only a few targets in a certain layer can pass the high-confidence screening, it means that the layer is not suitable as a unified benchmark. The reason for considering both indicators simultaneously is that a scale layer with only clear targets but too few targets may miss some real point light sources; while a scale layer with many targets but heavy noise may easily mistake noise hotspots for real light sources.

[0041] In practical applications, the selection criteria for the seed layer are as follows: among all scale layers, the scale layer with the highest comprehensive suitability index of high-frequency energy purity and identification stability is selected as the seed layer. High-frequency energy purity represents the ratio of the average coefficient amplitude of candidate spots in the layer to the average noise variance of the layer, while identification stability represents the number of objects in the layer whose confidence scores exceed a preset high threshold. In specific implementation, candidate spot detection is first performed on each scale layer, and then the average response strength of the decomposition coefficients in the candidate spot region is statistically analyzed. Simultaneously, the noise fluctuation level is estimated from the background region far from the candidate spots, thereby determining the high-frequency energy purity of the layer. Subsequently, multi-dimensional feature verification is performed on each candidate object in the layer, such as combining brightness concentration, local symmetry, connected region size stability, and consistency across neighboring scale layers, to output the confidence score of each candidate object. The system pre-sets a high confidence threshold; only candidate objects exceeding this threshold are included in the identification stability calculation. The "preset high threshold" here can be obtained through debugging with historical samples. For example, real point light sources and false spot samples are manually labeled from multiple sets of nighttime drone images, and the false detection rate and false negative rate under different thresholds are repeatedly tested. The threshold that balances the preservation of real light sources and the suppression of false detections is selected as the system operating parameter. To avoid excessive differences in the value range of indicators at different scales, before actually comparing the comprehensive suitability index, the high-frequency energy purity and recognition stability can be normalized to the same dimension, and then fused according to the preset weights. The weights can be determined through offline testing according to the application focus. If the scene emphasizes noise resistance, the weight of high-frequency energy purity can be increased; if the scene emphasizes complete recognition, the weight of recognition stability can be increased.

[0042] This specific implementation plan clarifies the calculation dimensions of high-frequency energy purity and identification stability, as well as the fusion method of the comprehensive suitability index. For example, after constructing a scale space containing multiple high-frequency detail layers, the algorithm evaluates each layer. For the i-th layer, the average amplitude of the decomposition coefficients within all candidate spot regions in this layer is first calculated, and the average noise variance of the non-spot background region in this layer is estimated. The high-frequency energy purity index is obtained by dividing the two. Simultaneously, the number of candidate objects in this layer with a normalized confidence score exceeding 0.8 (a preset high threshold) after multi-dimensional feature verification is counted and used as the identification stability index. Subsequently, these two indices are normalized separately and weighted according to preset weights (e.g., each accounting for 50%) to obtain the comprehensive suitability index for each layer. Comparison reveals that a specific layer has the highest comprehensive suitability index, indicating that this layer retains sufficient light source energy while having extremely low noise interference. Therefore, the algorithm automatically locks this layer as the seed layer for subsequent initial protection marker matrix generation. To make this example more actionable, the methods for obtaining candidate spots and background regions can be further explained: candidate spots can be obtained through local peak detection and connected component filtering, while background regions can be selected from dark background locations at a certain distance from the candidate spot boundaries to avoid the influence of halo trailing. Multidimensional feature verification can be implemented using rule-based judgment methods. For example, a higher confidence level is assigned only when a candidate object simultaneously meets the following conditions: its brightness is higher than the background, its region size falls within the allowable range, its central response is concentrated, and a corresponding response exists in adjacent scale layers. In this way, the comprehensive suitability index is no longer an abstract score, but is gradually derived from a directly operable detection, filtering, and statistical process.

[0043] The working principle of this specific implementation scheme lies in the objective measurement of the signal-to-noise ratio and target integrity of each scale layer by introducing two quantitative indicators: high-frequency energy purity and identification stability. High-frequency energy purity ensures that the selected layer has a clear light source response, while identification stability ensures that the structure of the real light source in the layer is not severely damaged by noise. The resulting technical effect is that it achieves fully automatic and adaptive selection of the seed layer, completely eliminating the blindness and limitations brought about by manually specifying a fixed scale layer. This allows the algorithm to always find the benchmark layer that best reflects the physical form of the light source when facing nighttime drone images with different shooting altitudes and different ambient brightness, thus providing the most reliable initial parameters for subsequent cross-layer boundary expansion and greatly enhancing the algorithm's scene adaptability. The underlying reason is that different flight altitudes will change the projection size of a single light source in the image, and different exposure strategies and ambient brightness will change the noise intensity and halo diffusion range, so fixed selection of a certain scale layer is not stable. By simultaneously evaluating whether the light source signal in the layer is sufficiently clean and whether the verifiable light source in the layer is sufficiently stable, the system can automatically avoid layers with excessive noise or edge diffusion. The seed layer ultimately selected is usually in a balanced position that is both "clear to see" and "accurate to recognize." Therefore, the centroid coordinates, initial core strong protection region radius, and initial peripheral weak protection region radius generated based on this seed layer are more reliable, and the protection results after cross-layer transfer are more consistent.

[0044] After identifying the seed layer and the actual point light source, the key to preserving the light source details lies in scientifically delineating the protection area and implementing differentiated protection strategies. Different types of light sources (such as strong streetlights and weak traffic lights) have different degrees of halo diffusion. Using a fixed protection radius will result in the halo of strong light sources being truncated or noise remaining around weak light sources. In order to accurately match the actual physical form of different light sources and implement refined control over the core and periphery, a specific scheme for dynamically delineating areas and hierarchical protection is proposed. The "halo area ratio characteristic" mentioned here is a characteristic quantity used to characterize the relationship between the bright core of a real point light source and the surrounding visible light diffusion range. Its essence reflects the diffusion strength of the light source under the current imaging conditions. The "core strong protection area" corresponds to the most important and concentrated bright core of the light source and its adjacent high-contrast edge, while the "peripheral weak protection area" corresponds to the halo band that gradually decays outside the bright core. Since these two areas perform different visual functions, the former emphasizes sharpness and main brightness preservation, while the latter emphasizes gradual transition and noise balance. Therefore, different protection strategies are required.

[0045] In a preferred embodiment of this application, for the identified real point light source, a core strong protection region and a surrounding weak protection region are delineated based on the centroid of the real point light source. Specifically, this includes: using the centroid coordinates of the real point light source as the origin, and combining the halo area ratio characteristics of the real point light source, dynamically calculating the radius of the core strong protection region and the radius of the surrounding weak protection region. In practice, the local bright spot region corresponding to the real point light source can be extracted from the seed layer first. Based on the brightness distribution, the central bright part is identified as the bright core region, and then the continuously decaying visible halo region around it is identified. The halo area ratio characteristics are obtained by statistically analyzing the area relationship between the two. If the characteristic shows that the bright core is small and the peripheral diffusion range is large, it indicates that the light source has a more obvious outward halo, and the radius of the surrounding weak protection region should be appropriately increased; if the characteristic shows that the bright core accounts for a large proportion and the peripheral diffusion is small, the radius of the surrounding weak protection region can be appropriately reduced. The radius of the core strong protection region usually surrounds the bright core and covers the main energy concentration area around the centroid, while the radius of the surrounding weak protection region extends outward to enclose the halo attenuation zone. Image data used to obtain halo area ratio features can be directly derived from the decomposition results of the seed layer, or combined with the local brightness distribution of the original night scene image for auxiliary confirmation to improve the fit of the boundary delineation. For the decomposition coefficients within the core strongly protected region, an absolute protection instruction is executed to preserve the original values, ensuring high contrast and sharpness of the light source center after reconstruction. This "absolute protection instruction" can be understood as not performing any shrinkage or weakening on the coefficients in this region during the filtering stage, and not participating in the ordinary background noise suppression process; instead, the original coefficients before decomposition are directly used in subsequent reconstruction. In implementation, a dedicated identifier bit can be written for the core strongly protected region in a dedicated protection marker matrix. The subsequent filtering module skips the threshold processing step for that pixel after reading this identifier bit. This ensures that the brightest and clearest structure at the center of the light source is not accidentally damaged by the denoising process, avoiding problems such as center graying, soft edges, or bright core collapse. For the decomposition coefficients within the surrounding weakly protected region, a protective smoothing threshold is calculated, and coefficients exceeding this threshold are shrunk towards zero. In practice, a protective smoothing threshold can be determined within the weakly protected area surrounding each real point light source, taking into account local noise variance, local brightness gradient, and the frequency band characteristics of the current scale layer. If a location is close to the core strong protected area and the brightness transition is obvious, the threshold can be appropriately relaxed to preserve more halo details; if a location is close to the outer edge of the surrounding weakly protected area and the noise proportion increases, the threshold can be relatively tightened to improve background cleansing capabilities. When performing zero-to-zero contraction, only the decomposition coefficients exceeding the protective smoothing threshold are compressed to a limited extent, while the decomposition coefficients below the threshold remain unchanged, thus preserving the true gradation of the halo edge while suppressing random noise and anomalous spikes.

[0046] This specific implementation plan details the extraction method of the halo area ratio feature, as well as the specific application of the absolute protection command and the protective smoothing threshold. For example, in the seed layer, for a identified real point light source, the algorithm first calculates the ratio of its bright core area to the overall visible halo area, i.e., the halo area ratio feature. If this ratio is small, it indicates a wide halo diffusion range, and the algorithm dynamically assigns a larger radius to the surrounding weak protection region; conversely, it assigns a smaller radius. After defining the region, corresponding protection commands are generated: for pixels falling within the core strong protection region, the shrinkage threshold of their decomposition coefficients is forcibly set to infinity, or directly marked as skipping filtering, thus executing the absolute protection command; for pixels falling within the surrounding weak protection region, the algorithm calculates a suitable protective smoothing threshold based on the local noise variance, and uses a soft thresholding function to moderately shrink the decomposition coefficients exceeding this threshold. Furthermore, to more clearly illustrate the extraction process of the halo area ratio feature, brightness stratification can be performed within the local neighborhood of the real point light source: the central region with the highest brightness and strongest connectivity is designated as the high-brightness core region, while the surrounding continuous region with gradually decreasing brightness is designated as the overall visible halo region. If, under the same acquisition conditions, the central region of a strong streetlight is small but its peripheral trailing range is large, then its surrounding weak protection zone needs to cover a greater distance to prevent the natural halo from being prematurely truncated during denoising. Conversely, if a weak signal light has a compact bright spot and a limited diffusion range, its surrounding weak protection zone should not be set too large; otherwise, noise from the surrounding dark background may be mistakenly included in the protection range, affecting the adequacy of denoising. Therefore, dynamically adjusting the protection radius using the halo area ratio feature can more closely approximate the actual imaging pattern.

[0047] The working principle of this specific implementation scheme lies in directly linking the physical diffusion characteristics of the light source (halo area ratio) with the delineation of the protection boundary, achieving customized area division for each light. Simultaneously, a stepped filtering strategy is implemented in different areas, with the core area fully preserved to maintain sharpness, while the peripheral areas are smoothed to suppress noise. The resulting technical effect is that it accurately matches the actual physical form of different types and brightness light sources, avoiding the halo cutoff of strong light sources or the residual noise around weak light sources caused by using a fixed protection radius. In the final reconstructed nighttime image, the real point light source not only maintains extremely high contrast and sharpness at its center, but its peripheral halo also exhibits an extremely natural and delicate gradual attenuation effect, perfectly balancing the dual needs of detail preservation and background noise reduction. Further analysis reveals that the core, strongly protected area requires absolute protection because it contains the main brightness peaks that determine the visual recognizability of the light source and the most crucial edge information. If this area is shrunk, the reconstructed image will directly exhibit insufficient brightness and blurred outlines. The surrounding, weakly protected area uses a protective smoothing threshold because this area contains both genuine halos and significant noise. Without filtering, the background will appear dirty; with excessive filtering, the halo will appear harsh or broken. By processing these two types of areas separately, the system achieves a stable balance between "no damage to the center, no breakage of the edges, and no background noise," making it particularly suitable for nighttime drone images where high-brightness light sources account for a large proportion and the background noise is significant in dark areas.

[0048] In the initial stage of multi-scale denoising, accurately extracting candidate real light source signals from complex background noise is a prerequisite for the implementation of all subsequent protection mechanisms. As mentioned earlier, identifying cross-scale stable high-response blobs in scale space is described. To further clarify the specific implementation of this identification process and ensure the robustness and accuracy of the extraction results, in a preferred embodiment of this application, identifying cross-scale stable high-response blobs in scale space includes: traversing the high-frequency detail feature maps of each scale layer; for each feature map, calculating an adaptive threshold based on the absolute median difference; marking pixels with coefficient absolute values ​​greater than the adaptive threshold as candidate points and merging them into candidate high-response blobs using a connected component algorithm, and recording their centroid coordinates; using the finest scale layer as a reference, projecting the centroid coordinates of its blobs onto the adjacent coarse scale layer, searching for matching blobs within a preset search tolerance radius of the projected coordinates, and if a match is found, establishing a cross-scale bidirectional link pointer, repeating the projection link process to traverse all scale layers; based on the cross-scale bidirectional link pointer, counting the number of existence layers of each object in scale space, and marking objects whose proportion of existence layers to the total number of decomposition layers meets a preset persistence ratio threshold as cross-scale stable high-response blobs.

[0049] The specific implementation scheme described above details the extraction and linking logic for stable high-response blobs across scales. In practice, firstly, for the high-frequency detail feature map at each scale layer, an adaptive threshold is calculated using the absolute median. The absolute median is a statistical measure with extremely strong robustness to outliers. In the high-frequency feature map of nighttime images, the coefficients of most background pixels approach zero, while light source edges and strong noise exhibit extremely high responses. If the threshold is calculated using conventional mean or variance, it is easily inflated by these bright outliers, causing the threshold to fail; however, the adaptive threshold calculated based on the absolute median can stably reflect the true fluctuation level of background noise. Subsequently, pixels exceeding this threshold are extracted, and spatially adjacent pixels are merged into independent candidate high-response blobs using a connected component algorithm, and their centroid coordinates are calculated.

[0050] After extracting speckles in a single layer, the system performs a cross-scale linking operation. Since the finest-scale layer (i.e., the highest-frequency layer) retains the sharpest edge information of the image and its speckle localization is most accurate, it is used as the projection reference. The centroid coordinates of the speckles in the finest-scale layer are projected onto the adjacent coarser-scale layer (lower-frequency layer). Considering that the light source undergoes physical diffusion in the low-frequency layer, and that multi-scale transformation may introduce slight phase shifts, the coordinates are not required to perfectly coincide after projection. Instead, a matching speckle is searched for within a preset search tolerance radius. If a match is found, a bidirectional cross-scale link pointer is established between the two specks to clarify their inter-layer correspondence. This projection and linking process is traversed layer by layer downwards until the entire scale space is covered. Finally, the system counts the number of layers in which each speckle object continuously exists throughout the entire scale space along the established bidirectional link pointers. If the proportion of this number of existing layers to the total number of decomposed layers reaches a preset persistence threshold (e.g., requiring a response in more than 70% of scale layers), it is officially marked as a cross-scale stable high-response speckle.

[0051] By employing the aforementioned technical solutions, and utilizing absolute median difference adaptive thresholding and connected component analysis, background noise can be effectively removed and high-response regions can be accurately located within a single layer. Furthermore, the projection and bidirectional linking mechanism based on the finest-scale layer benchmark establishes a rigorous spatiotemporal correspondence between multiple layers. The resulting technical effect is that it can extremely efficiently eliminate random patch noise that only occasionally appears in individual frequency bands, filtering out light source signals that truly possess cross-scale continuous physical structures. This provides a high-quality, high-purity set of candidate targets for subsequent feature discrimination, significantly improving the anti-interference capability of target recognition in complex night scenes.

[0052] After identifying stable, high-response spots across scales, although most random noise has been eliminated, there may still be some persistent interferences in urban nighttime scenes that resemble the shape of real light sources (such as strong reflective spots, sensor dead pixels, etc.). The identification logic based on multi-dimensional features has been briefly described above. To further quantify this identification process and achieve accurate automated decision-making, this application identifies the real point light source to be protected based on multi-dimensional features of optical morphology, energy, and spatial context. Specifically, it is a multi-level serial verification and weighted scoring process, including: the first level for cross-scale stable high-response spots, evaluating the fitting error of the local brightness distribution to the two-dimensional Gaussian point spread function and the halo area ratio, and calculating the optical morphology score; the second level extracts the response amplitude of cross-scale stable high-response spots at each scale level and forms a cross-scale energy decay sequence according to the scale order, evaluates the smoothness variance of the sequence, and calculates the energy decay score; the third level back-projects the pixel coordinates of cross-scale stable high-response spots to the world coordinate system, evaluates the collinearity error of the world coordinate system, and calculates the spatial context score; the scores of the three levels are weighted and summed according to preset weights to obtain a normalized confidence score, which is compared with a preset threshold to determine the real point light source.

[0053] This specific implementation scheme constructs a multi-level cascaded verification system from the inside out and from the local to the global. The first level of verification focuses on the morphological evaluation of microscopic physical optics. Ideal point light sources in optical imaging systems typically follow a two-dimensional Gaussian point diffusion distribution. Therefore, by calculating the fitting error between the local brightness distribution of candidate spots and the standard two-dimensional Gaussian model, and combining this with the halo area ratio characteristic mentioned earlier, the regularity and naturalness of their morphology can be quantified, and thus mapped to an optical morphology score. The smaller the fitting error and the more the halo ratio conforms to the characteristics of a real light source, the higher the score.

[0054] The second-level verification focuses on evaluating the energy evolution pattern across frequency bands. In real light sources, under multi-scale decomposition, the energy transition from high to low frequencies is typically continuous and smooth; however, persistent noise or anomalous reflections often exhibit abrupt energy changes within a specific frequency band. Therefore, the system extracts the response amplitudes of candidate spots at each scale level, arranges them in hierarchical order to form a cross-scale energy attenuation sequence, and calculates the smoothness variance of this sequence. A smaller variance indicates smoother energy attenuation, resulting in a higher energy attenuation score.

[0055] In the third level of verification, the focus is on the logical evaluation of macroscopic geospatial conditions. As mentioned earlier, the image pixel coordinates are back-projected to the world coordinate system using the UAV pose and camera intrinsic parameters. Based on this, the system further calculates the collinear arrangement error or regular distribution error of these candidate targets in three-dimensional space. If multiple targets exhibit a collinear arrangement that highly matches the road direction in the world coordinate system, their error is minimal, and they are assigned a very high spatial context score; conversely, if they are isolated and disordered, the score is lower.

[0056] Finally, the system weights and sums the scores from the three levels according to preset weights, outputting a normalized confidence score. Only when this score exceeds a preset threshold is the candidate spot finally confirmed as a real point light source to be protected.

[0057] The aforementioned technical solution organically combines microscopic optical imaging mechanisms, mesoscopic frequency band energy patterns, and macroscopic spatial geographic distribution to form a rigorous quantitative scoring mechanism. Multi-level cascaded verification not only avoids misjudgments easily caused by single features but also achieves complementary advantages of multi-dimensional information through weighted summation. The resulting technical effect is a significant reduction in the false detection and false negative rates of landmark light sources in complex urban night scene images. Even in extreme low-light environments with severe background light pollution and extremely complex noise patterns, it can still accurately and stably identify real point light sources requiring protection, providing the most solid data support for subsequent cross-layer boundary expansion and differentiated denoising.

[0058] Based on the above multi-scale image denoising methods, such as Figure 3 As shown, this application also proposes a multi-scale image denoising system 100 for unmanned aerial vehicles, comprising: The luminance channel decomposition module 10 is used to decompose the luminance channel of the nighttime low-light image acquired by the UAV through non-subsampling transformation and construct a scale space containing multiple scale layers. The real point source identification module 20 is used to identify cross-scale stable high-response spots in scale space and to identify real point sources to be protected based on multi-dimensional features of energy, optical morphology and spatial context. The initial protection mark matrix generation module 30 is used to delineate the core strong protection area and the surrounding weak protection area based on the centroid of the identified real point light source, and generate the seed layer initial protection mark matrix. The inter-layer transfer and verification module 40 is used to take the seed layer as the first scale layer, decompose the initial protection mark matrix into a structured instruction package, and transfer it to the second scale layer. The instruction package contains the initial boundary parameters of the core strong protection region. The second scale layer calculates the boundary expansion coefficient based on the layer difference between itself and the first scale layer. The initial boundary parameters are adjusted using the boundary expansion coefficient to obtain the target boundary parameters adapted to the second scale layer. Based on the target boundary parameters, a dedicated protection mark matrix is ​​regenerated in the second scale layer to guide the denoising operation of the second scale layer.

[0059] Through the aforementioned technical solution, this UAV multi-scale image denoising system highly integrates and modularizes complex steps such as image decomposition, feature identification, region delineation, and cross-layer parameter transfer. The modules collaborate efficiently through standardized data interfaces, ensuring the continuity of the processing chain while significantly reducing system operating costs. This system can automatically and accurately perform denoising tasks on low-light nighttime images. While powerfully suppressing background patch noise, it perfectly preserves the high-contrast brightness cores and natural gradient halos of urban landmark light sources, completely solving the technical problems of blurred light sources and fragmented edges that traditional denoising systems often cause. This significantly improves the visual quality of UAV nighttime inspection images and the reliability of subsequent intelligent analysis.

[0060] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for multi-scale image denoising of unmanned aerial vehicles (UAVs), characterized in that, include: The brightness channel of low-light nighttime images acquired by UAVs is decomposed using non-subsampling transformation to construct a scale space containing multiple scale layers. In the aforementioned scale space, cross-scale stable high-response spots are identified, and the real point light sources to be protected are identified based on multi-dimensional features of optical morphology, energy, and spatial context. For the identified real point light sources, the core strong protection area and the surrounding weak protection area are delineated based on the centroid of the real point light source, and the initial protection mark matrix of the seed layer is generated. Using the seed layer as the first scale layer, the initial protection mark matrix is ​​decomposed into a structured instruction package and passed to the second scale layer. The structured instruction package contains the initial boundary parameters of the core strong protection region. The second-scale layer calculates the boundary expansion coefficient based on its level difference with the first-scale layer; the initial boundary parameters are adjusted using the boundary expansion coefficient to obtain target boundary parameters that fit the second-scale layer; and based on the target boundary parameters, a dedicated protection marker matrix is ​​regenerated in the second-scale layer to guide the denoising operation of the second-scale layer.

2. The method according to claim 1, characterized in that, The steps for calculating the boundary spread coefficient of the second-scale layer based on its layer difference with the first-scale layer include: The second-scale layer calculates the boundary expansion coefficient using a monotonically increasing function based on the difference in layer depth between itself and the first-scale layer; wherein the difference in layer depth is the difference between the layer index of the second-scale layer and the layer index of the first-scale layer; the boundary expansion coefficient increases with the increase of the difference in layer depth, so that the boundary range of the protected area monotonically expands as the frequency band decreases.

3. The method according to claim 2, characterized in that, The method includes: When adjusting the initial boundary parameters using the boundary expansion coefficient, the initial core protection radius and the initial peripheral protection radius are multiplied by the boundary expansion coefficient to obtain the target core protection radius and the target peripheral protection radius. The initial core protection radius corresponds to the scale of the core strong protection area in the first scale layer, and the initial peripheral protection radius corresponds to the scale of the peripheral weak protection area in the first scale layer. The values ​​of the target core protection radius and the target peripheral protection radius both increase with the increase of the layer depth difference.

4. The method according to claim 1, characterized in that, The method includes: When regenerating the dedicated protection marker matrix at the second scale layer, the existence of boundary conflict states is detected pixel by pixel; the boundary conflict state refers to a pixel coordinate falling into the weak protection area of ​​two or more different real point light sources at the same time, and the pixel falling into the overlapping area is the conflicting pixel. If a boundary conflict exists, the mild smoothing protection threshold corresponding to each light source that caused the conflict is extracted, and the controlled smoothing threshold of the conflicting pixel is set to the maximum value among the mild smoothing protection thresholds corresponding to all light sources. The value is used to perform controlled smoothing operation on the decomposition coefficients of the conflicting pixel. The mild smoothing protection threshold is a threshold used to control the noise suppression intensity in the surrounding weak protection area, and only coefficients exceeding this threshold are subjected to mild contraction toward zero.

5. The method according to claim 1, characterized in that, The structured instruction package includes: the hierarchical index of the first scale layer, the centroid coordinates of each confirmed real point light source in the first scale layer, the initial core strong protection region radius, and the initial peripheral weak protection region radius. The structured instruction packet, as a lightweight meta-data packet, is broadcast to all other target scale layers in the scale space through a cross-layer communication mechanism.

6. The method according to claim 1, characterized in that, The selection criteria for the seed layer are as follows: Among all scale layers, the scale layer with the highest comprehensive suitability index of high-frequency energy purity and identification stability is selected as the seed layer; wherein, the high-frequency energy purity represents the ratio of the average coefficient amplitude of candidate spots in the layer to the average noise variance of the layer, and the identification stability represents the number of objects in the layer whose confidence score exceeds a preset high threshold.

7. The method according to claim 1, characterized in that, The process involves defining a core strong protection zone and a surrounding weak protection zone for the identified real point light sources, using the centroid of the real point light source as a reference. Specifically, this includes: Using the centroid coordinates of the real point light source as the origin, and combining the halo area ratio characteristics of the real point light source, the radius of the core strong protection area and the radius of the surrounding weak protection area are dynamically calculated. For the decomposition coefficients within the core strong protection area, an absolute protection command is executed to retain the original values, ensuring high contrast and sharpness at the center of the light source after reconstruction; For the decomposition coefficients within the surrounding weakly protected area, calculate the protective smoothing threshold, and perform a shrinkage towards zero on coefficients exceeding the protective smoothing threshold.

8. The method according to claim 1, characterized in that, Identifying cross-scale stable high-response spots in the scale space includes: Traverse the high-frequency detail feature maps of each scale layer, and for each feature map, calculate an adaptive threshold based on the absolute median difference; Pixels whose absolute value of the coefficient is greater than the adaptive threshold are marked as candidate points and merged into candidate high-response spots by a connected component algorithm, and their centroid coordinates are recorded. Using the finest scale layer as a reference, the centroid coordinates of its spots are projected onto the adjacent coarse scale layer. Matching spots are found within the preset search tolerance radius of the projected coordinates. If a match is found, a bidirectional link pointer across scales is established. The projection link process is repeated to traverse all scale layers. Based on cross-scale bidirectional link pointers, the number of existence layers of each object in the scale space is counted. Objects whose existence layer ratio meets the preset persistence ratio threshold are marked as cross-scale stable high-response spots.

9. The method according to claim 1, characterized in that, The identification of the actual point light source to be protected based on multi-dimensional features of optical morphology, energy, and spatial context specifically involves a multi-level cascaded verification and weighted scoring process, including: The first level evaluates the fitting error between the local brightness distribution and the two-dimensional Gaussian point spread function, as well as the halo area ratio, for the cross-scale stable high-response spot, and calculates the optical morphology score. The second stage extracts the response amplitude of the cross-scale stable high-response spot at each scale layer and forms a cross-scale energy decay sequence in scale order. The smoothness variance of the sequence is evaluated and the energy decay score is calculated. The third level back-projects the pixel coordinates of the cross-scale stable high-response spots to the world coordinate system, evaluates their collinear alignment error in the world coordinate system, and calculates the spatial context score. The three levels of scores are weighted and summed according to preset weights to obtain a normalized confidence score, which is then compared with a preset threshold to determine the true point light source.

10. A multi-scale image denoising system for unmanned aerial vehicles (UAVs), characterized in that, include: The luminance channel decomposition module is used to decompose the luminance channels of low-light nighttime images acquired by UAVs through non-subsampling transformation, and construct a scale space containing multiple scale layers. The real point light source identification module is used to identify cross-scale stable high-response spots in the scale space and to identify the real point light sources to be protected based on multi-dimensional features of energy, optical morphology and spatial context. The initial protection marker matrix generation module is used to delineate the core strong protection area and the surrounding weak protection area based on the centroid of the identified real point light source, and generate the seed layer initial protection marker matrix. The inter-layer transfer and verification module is used to take the seed layer as the first scale layer, decompose the initial protection marker matrix into a structured instruction package, and transfer it to the second scale layer. The structured instruction package contains the initial boundary parameters of the core strong protection region. The second scale layer calculates the boundary expansion coefficient based on its level difference with the first scale layer. The initial boundary parameters are adjusted using the boundary expansion coefficient to obtain target boundary parameters adapted to the second scale layer. Based on the target boundary parameters, a dedicated protection marker matrix is ​​regenerated in the second scale layer to guide the denoising operation of the second scale layer.

Images