Unmanned aerial vehicle imaging adaptive enhancement method for low-illumination environment

Abstract

The present application relates to the technical field of unmanned aerial vehicle image processing, in particular to a low-illumination environment-oriented unmanned aerial vehicle imaging adaptive enhancement method, comprising: dividing the collected river water quality inspection images into regions to obtain each water body region image; extracting the light attenuation, near-water surface scattering and water surface reflection degradation parameters of each water body region image, generating positive and negative polarity adjustment features and performing polarity fusion to obtain region enhancement control parameters and enhance each water body region image; obtaining the water body texture and edge gradient change amount by comparing the images before and after enhancement, and accordingly feeding back and correcting the enhancement control parameters to realize re-enhancement and splicing output. Through the targeted polarity feature extraction and fusion adjustment of the multi-factor coupled degradation parameters, the local adaptability of different water body regions is improved; combined with the double-index feedback correction and limiting mechanism, the loss of water body texture and edge gradient features is reduced while improving the image visibility.

Description

Technical Field

[0001] This invention relates to the field of UAV image processing technology, specifically to an adaptive enhancement method for UAV imaging in low-light environments. Background Technology

[0002] With the development of drone platforms and airborne imaging equipment, drones have been applied to the inspection of water bodies such as rivers, lakes, and reservoirs, acquiring large-scale water images and providing data support for water quality analysis and anomaly monitoring. However, in early morning, high humidity, and foggy environments, the acquired images are easily affected by insufficient light and fog scattering, resulting in problems such as reduced brightness, loss of detail, and decreased contrast.

[0003] Compared to land images, water areas possess unique imaging characteristics such as dynamic reflection, complex light propagation, and weak texture information. In low-light and foggy environments, light attenuation and fog scattering interact, while water surface reflection interference is superimposed, resulting in water images exhibiting a multi-factor coupled effect of light attenuation, near-water surface scattering, and water surface reflection. Furthermore, due to the weak texture information in water areas, differences in the spatial distribution of these factors can easily lead to deviations in the extraction of light attenuation parameters, near-water surface scattering parameters, and water surface reflection degradation parameters.

[0004] Existing low-light enhancement and dehazing methods typically correct for single factors and employ a uniform enhancement strategy, making it difficult to adapt to differences in light attenuation, fog scattering, and water surface reflection in different water areas during UAV water quality inspections. Current methods often adjust image brightness or overall contrast during enhancement, easily altering the original weak contrast structure of the water body and changing two key imaging features for water quality analysis: water texture and edge gradients. Based on this analysis, during UAV water quality inspections, it is necessary to improve the visibility of water images in low-light and foggy environments while avoiding the loss of water texture and edge gradients during enhancement processing, thus meeting the requirement for preserving detailed features in water quality analysis.

[0005] To address this, an adaptive enhancement method for UAV imaging in low-light environments is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive enhancement method for UAV imaging in low-light environments. By extracting light attenuation parameters, near-water surface scattering parameters, and water surface reflection degradation parameters respectively, positive and negative polarity adjustment features are generated for constraint fusion. An adaptive feedback correction and amplitude limiting mechanism is constructed by combining water texture changes and edge gradient changes. This improves the problem that existing methods, which use a uniform enhancement strategy for a single factor, are difficult to adapt to the differences in different water areas. It also alleviates the risk that the original weak contrast structure of the water body can be easily altered when enhancement is adjusted or parameter extraction deviations occur, thereby causing damage or loss of water texture and edge gradient information used for water quality analysis.

[0007] To achieve the above objectives, the present invention provides the following technical solution: Adaptive augmentation methods for UAV imaging in low-light environments include: By collecting river water quality inspection images using drones, the spatial direction of the river channel and the texture of the water-land boundary are extracted. Based on the spatial direction of the river channel and the texture of the water-land boundary, the river water quality inspection images are divided into regions to obtain images of each water body region. Extract illumination attenuation parameters, near-water surface scattering parameters, and water surface reflection degradation parameters from images of each water body area; generate positive polarity illumination modulation features based on illumination attenuation parameters, negative polarity scattering modulation features based on near-water surface scattering parameters, and negative polarity reflection modulation features based on water surface reflection degradation parameters. Polarity fusion is performed on positive polarity illumination modulation features, negative polarity scattering modulation features, and negative polarity reflection modulation features to generate regional enhancement control parameters; based on the regional enhancement control parameters, the images of each water body region are enhanced, and the changes in water texture and edge gradient are obtained by comparing the images of each water body region before and after enhancement. The regional enhancement control parameters are corrected based on the changes in water texture and edge gradient. The images of each water body region are then re-enhanced based on the corrected regional enhancement control parameters to obtain optimized water body region images. The optimized water body region images are then stitched together by region to output optimized river water quality inspection images.

[0008] Preferably, the process of dividing the river water quality inspection image into regions based on the spatial direction of the river channel and the texture of the water-land boundary includes: extracting water texture features and edge gradient features from the river water quality inspection image; determining the texture of the water-land boundary based on the water texture features and edge gradient features; determining the water-land boundary region based on the texture of the water-land boundary; determining the water body region range based on the water-land boundary region, and dividing the river water quality inspection image along the spatial direction of the river channel according to the water body region range to obtain each initial region image; removing isolated pixel connected blocks and non-water body abnormal pixel connected blocks from each initial region image to obtain each water body region image.

[0009] Preferably, the process of removing isolated pixel connected blocks and non-water body abnormal pixel connected blocks from each initial region image includes: identifying pixel connected blocks in each initial region image based on the connectivity distance between adjacent pixels in each initial region image; obtaining the connectivity area, spatial location, and color features of the pixel connected blocks; removing pixel connected blocks with a connectivity area less than a preset area threshold as isolated pixel connected blocks; and removing pixel connected blocks whose spatial location is within the water-land boundary area, as well as pixel connected blocks whose color features deviate from a preset water body color range, as non-water body abnormal pixel connected blocks.

[0010] Preferably, the specific process of extracting the illumination attenuation parameters, near-water surface scattering parameters, and water surface reflection degradation parameters of the images of each water body region includes: acquiring the overall brightness distribution characteristics, local contrast attenuation characteristics, water surface transmission change characteristics, high-brightness anomaly distribution characteristics, and directional reflection characteristics of the images of each water body region; calculating the average brightness value based on the overall brightness distribution characteristics, and determining the illumination attenuation parameters based on the deviation of the average brightness value from the preset reference brightness value; determining the local contrast characterization quantity based on the local contrast attenuation characteristics, determining the transmittance spatial change characterization quantity based on the water surface transmission change characteristics, and weighting the local contrast characterization quantity and the transmittance spatial change characterization quantity according to the preset scattering evaluation weight to determine the near-water surface scattering parameters; calculating the abnormal high-brightness pixel density based on the high-brightness anomaly distribution characteristics, determining the reflection direction intensity based on the directional reflection characteristics, and multiplying the abnormal high-brightness pixel density by the reflection direction intensity to determine the water surface reflection degradation parameters.

[0011] Preferably, the polarity fusion of positive polarity illumination modulation features, negative polarity scattering modulation features, and negative polarity reflection modulation features to generate region enhancement control parameters includes: calculating the regional feature amplitudes of the positive polarity illumination modulation features, negative polarity scattering modulation features, and negative polarity reflection modulation features in the images of each water body region; normalizing each regional feature amplitude to obtain illumination weights, scattering weights, and reflection weights respectively; multiplying the positive polarity illumination modulation features by the illumination weights, the negative polarity scattering modulation features by the scattering weights, and the negative polarity reflection modulation features by the reflection weights, and algebraically superimposing the product results to obtain preliminary fusion parameters; and generating region enhancement control parameters based on the preliminary fusion parameters.

[0012] Preferably, the specific process of generating region enhancement control parameters based on preliminary fusion parameters includes: determining the parameter change gradient based on the difference between preliminary fusion parameters of spatially adjacent water body region images; when the parameter change gradient is greater than a preset gradient threshold, applying spatial continuity constraints to the preliminary fusion parameters to obtain region enhancement control parameters; when the parameter change gradient is not greater than the preset gradient threshold, using the preliminary fusion parameters as region enhancement control parameters.

[0013] Preferably, the specific process of correcting the region enhancement control parameters based on the water texture change and edge gradient change includes: extracting water texture features and edge gradient features from the images of each water region before and after enhancement; calculating the difference between the water texture features before and after enhancement to obtain the water texture change; calculating the difference between the edge gradient features before and after enhancement to obtain the edge gradient change; determining the texture protection component based on the deviation of the water texture change from a preset texture continuity threshold; determining the boundary perturbation degree based on the edge gradient change; and determining the boundary protection component based on the deviation of the boundary perturbation degree from preset first thresholds and preset second thresholds; wherein, the preset... The first threshold is less than the preset second threshold. When the boundary disturbance degree is not greater than the preset first threshold, the boundary protection component is the difference between the preset first threshold and the boundary disturbance degree. When the boundary disturbance degree is not less than the preset second threshold, the boundary protection component is the difference between the preset second threshold and the boundary disturbance degree. When the boundary disturbance degree is greater than the preset first threshold and less than the preset second threshold, the boundary protection component is zero. The texture protection component and the boundary protection component are weighted and summed according to the preset correction weight to obtain the parameter correction amount. The parameter correction amount is algebraically added to the region enhancement control parameters to obtain the initial addition result. The corrected region enhancement control parameters are generated based on the initial addition result.

[0014] Preferably, the specific process of generating the corrected region enhancement control parameters based on the initial addition result includes: constraining the initial addition result according to a preset enhancement parameter constraint range to obtain constrained region enhancement control parameters; enhancing the water area image based on the constrained region enhancement control parameters to obtain a constrained enhanced image; extracting the real-time feature matrix corresponding to the constrained enhanced image and comparing the real-time feature matrix with a pre-stored reference water body feature matrix to obtain feature difference results; when the feature difference results are not greater than a preset difference upper limit, using the constrained region enhancement control parameters as the corrected region enhancement control parameters; when the feature difference results are greater than the preset difference upper limit, using the region enhancement control parameters before algebraic addition as a benchmark, truncating the constrained region enhancement control parameters and using the truncated region enhancement control parameters as the corrected region enhancement control parameters.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention extracts light attenuation parameters, near-water surface scattering parameters, and water surface reflection degradation parameters separately to generate corresponding positive or negative polarity adjustment features, which are then fused after normalization. This process achieves separate characterization and fusion adjustment of degradation features under the coupled effects of multiple factors such as light attenuation, near-water surface scattering, and water surface reflection. It improves upon existing low-light enhancement and defogging methods, which typically correct for a single factor and employ a uniform enhancement strategy, making it difficult to adapt to the differences in light attenuation, fog scattering, and water surface reflection in different water areas during UAV water quality inspections. Compared to globally uniform enhancement methods, this invention can adjust enhancement parameters according to the degradation characteristics of different water areas, improving the adaptability of water area enhancement processing under multiple factors and enhancing the visibility of water images in low-light environments.

[0016] 2. This invention compares water region images before and after enhancement to obtain changes in water texture and edge gradient. Based on these changes, a feedback correction mechanism is constructed to obtain corrected region enhancement control parameters for subsequent image enhancement processing. This process dynamically adjusts the enhancement parameters according to changes in water region features before and after enhancement, improving upon existing methods that typically adjust based on image brightness or overall contrast during enhancement. This can easily alter the original weak contrast structure of the water body, causing changes to water texture and edge gradient—two imaging features used for water quality analysis. This invention can improve image visibility in low-light and foggy environments while constraining and preserving water texture and edge gradient features, reducing the impact of enhancement processing on changes in detail features.

[0017] 3. In the parameter correction process, this invention introduces a preset enhancement parameter constraint range to constrain the initial summation result. It then compares the real-time feature matrix corresponding to the enhancement result with a pre-stored baseline water body feature matrix. Based on the comparison result, a limiting constraint of a preset ratio range is applied to obtain the region enhancement control parameters after multiple constraints. This process achieves joint constraints on the variation range of enhancement parameters and the distribution of enhanced color features, improving the problem that the extraction of light attenuation parameters, near-water surface scattering parameters, and water surface reflection degradation parameters is easily biased due to the weak texture information of the water body region and the spatial distribution differences of different factors. This invention can reduce the impact of parameter extraction bias on the enhancement result, enabling the water body image to maintain matching with the baseline water body characteristics while improving visibility, providing a more stable image data foundation for subsequent water quality analysis. Attached Figure Description

[0018] Figure 1 A flowchart of the UAV imaging adaptive enhancement method for low-light environments provided by the present invention; Figure 2A flowchart for water quality inspection image region division and anomaly removal provided by the present invention; Figure 3 The flowchart for multi-factor degradation parameter extraction and polarity fusion provided by this invention; Figure 4 The flowchart for the enhanced control parameter dual-index closed-loop feedback correction provided by the present invention; Figure 5 This is a schematic diagram of the water quality inspection image region division provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0020] Example 1: This embodiment describes the overall process of the method, which is applicable to the pre-flight parameter calibration and deployment stage of UAV water quality inspection missions, as well as the subsequent aerial online inspection operation stage in low light and foggy environments such as early morning.

[0021] Please see Figures 1 to 5 This invention provides an adaptive enhancement method for UAV imaging in low-light environments. The technical solution is as follows: by collecting river water quality inspection images by UAV, extracting the spatial direction of the river channel and the texture of the water-land boundary, and dividing the river water quality inspection images into regions based on the spatial direction of the river channel and the texture of the water-land boundary to obtain images of each water body region. Extract illumination attenuation parameters, near-water surface scattering parameters, and water surface reflection degradation parameters from images of each water body area; generate positive polarity illumination modulation features based on illumination attenuation parameters, negative polarity scattering modulation features based on near-water surface scattering parameters, and negative polarity reflection modulation features based on water surface reflection degradation parameters. Polarity fusion is performed on positive polarity illumination modulation features, negative polarity scattering modulation features, and negative polarity reflection modulation features to generate regional enhancement control parameters; based on the regional enhancement control parameters, the images of each water body region are enhanced, and the changes in water texture and edge gradient are obtained by comparing the images of each water body region before and after enhancement. The regional enhancement control parameters are corrected based on the changes in water texture and edge gradient. The images of each water body region are then re-enhanced based on the corrected regional enhancement control parameters to obtain optimized water body region images. The optimized water body region images are then stitched together by region to output optimized river water quality inspection images.

[0022] Furthermore, referring to Figure 2 and Figure 5The process of dividing river water quality inspection images into regions based on the spatial direction of the river channel and the texture of the water-land boundary includes: extracting water texture features and edge gradient features from the river water quality inspection images; determining the texture of the water-land boundary based on the water texture features and edge gradient features; determining the water-land boundary region based on the water-land boundary texture; determining the water body region range based on the water-land boundary region, and dividing the river water quality inspection images along the spatial direction of the river channel according to the water body region range to obtain each initial region image; removing isolated pixel connected blocks and non-water body abnormal pixel connected blocks from each initial region image to obtain each water body region image.

[0023] Specifically, the water quality inspection images collected by the drone during water quality inspection operations are used as the initial color images. The dimensions are Traverse the initial color image. All pixel coordinates Extract its , , The luminance components of the three color channels are calculated using a weighted average formula. The grayscale values ​​are calculated pixel by pixel, and the calculated grayscale values ​​are then sorted according to the original pixel coordinates. Spatial location mapping generates grayscale image matrix Meanwhile, the initial color image is retained. This is used in subsequent steps to extract color data from the target area. A local neighborhood window is created, and its size is set to balance the sufficiency of local texture statistics with the preservation of water and land edge resolution. To adapt to inspection images with different spatial resolutions, the window size parameter is set as follows. Introducing a texture window scaling factor associated with the image spatial size. Its preferred value range is 0.1% to 0.5%. Combined with image width... With height Using formula Calculations are performed to ensure that the window size dynamically adapts and is always an odd number, where, This is the floor function. This is the function for finding the minimum value. In this embodiment, a standard resolution image is used as an example; the specific value is determined through calculation. Set the total number of target quantization levels to [number]. To ensure the effectiveness of local information entropy statistics and to make the quantization level adaptive to the window size, the formula is adopted. Calculate (where (This is the floor function). In this embodiment, it corresponds to the aforementioned... The specific value after calculation is For grayscale image matrices any pixel in The corresponding quantization level index is calculated using the floor function mapping formula. : ; in, The range of values ​​is Traverse the grayscale image matrix Generate a quantization index matrix from all pixels. In the quantization index matrix, use any pixel... Centered on the local neighborhood window, the quantization levels are statistically analyzed. Frequency of occurrence Divide by the total number of pixels in the window The quantization levels were calculated. probability of occurrence The local entropy of the corresponding center pixel is calculated using the local entropy formula. : ; in, It is a very small constant, and in this embodiment it is taken as a value of Traverse the grayscale image matrix For all pixels, the calculated local entropy will be... Construct a local entropy matrix by arranging the corresponding spatial coordinates. The local entropy matrix is ​​normalized using the extremum normalization method. Normalization is performed. To avoid division by zero errors caused by flat inputs, piecewise function calculation is used. hour: ; when hour, And record that the current region is in a flat input state. and Representing the local entropy matrix respectively The maximum and minimum values ​​of the elements in the matrix are determined; the normalized local entropy matrix is ​​used as a two-dimensional matrix representing the water texture features. .

[0024] The bidirectional Sobel partial derivative operator is used in the grayscale image matrix Gradient information is extracted from the top. A structure with a size of [size missing] is constructed. Horizontal direction operator with vertical direction operator : ; Initialize a two-dimensional gradient empty matrix with the same spatial dimensions as the grayscale image matrix. Traverse the grayscale image matrix. All pixel coordinates except the boundary , will the operator and Align the center anchor point with the current pixel, and perform two-dimensional discrete convolution operations to compute the horizontal gradient components. and vertical gradient components The calculation formula is: ; ; in, and Operators and of OK, The matrix elements of the columns. Then, according to the formula... Calculate the gradient magnitude pixel by pixel, and then calculate each gradient magnitude. The gradient magnitude matrix is ​​constructed by filling the empty two-dimensional gradient matrix with its spatial coordinates. The gradient magnitude matrix is ​​then normalized using the same extremum normalization formula as the local entropy matrix, and the normalized matrix is ​​used as the two-dimensional feature matrix representing the edge gradient features. .

[0025] For the water texture feature matrix With edge gradient feature matrix Perform cross-response calculation. For any pixel coordinate point... Its cross-response value Calculated using the following formula: ; in, and Represent matrices respectively With matrix mid-coordinate point Matrix elements at the location; Represented by natural constant The natural logarithm function with base 0. Iterates through the pixel coordinates of the input image. The cross-response values ​​corresponding to each pixel coordinate point are arranged according to their spatial position to construct a two-dimensional cross-response feature matrix. .

[0026] The cross-response feature matrix is ​​analyzed using the Otsu's method. Perform adaptive threshold calculation. Specifically, set the total number of discrete feature levels to [value]. In this embodiment, it is preferred that... Extracting the matrix Minimum of characteristics With eigenvalue To avoid division-by-zero errors caused by flat inputs or a single input distribution, the following branching logic is used for handling: when When, for any coordinate point in the matrix The eigenvalues ​​at a given location are used to calculate their corresponding discrete feature level indices using a linear quantization formula. : ; in, This is a rounding function, feature-level index. The range of values ​​is Statistical analysis of each discrete feature level In the matrix Frequency of occurrence Divide by the total number of pixels in the entire image The feature levels are calculated. Corresponding probability of occurrence For any set feature-level threshold variable ( ), calculate the cumulative probability of the background class Cumulative probability with the target class When calculating the class mean, for or Threshold candidates are directly excluded from the candidate set without calculating their class mean and inter-class variance. For those that satisfy... and Valid candidate threshold Calculate the feature mean of the background class The feature mean of the target class Calculate the between-class variance. Traversal Solve for all possible values ​​such that the between-class variance is... When the maximum value is obtained The optimal segmentation threshold is calculated using the following inverse mapping formula. : ; when If no valid candidate threshold exists, determine that the current image is a flat input state and output a water mask matrix with all zeros. And terminate the subsequent segmentation calculation of the current region.

[0027] For normal branches with an effective threshold, the cross-response feature matrix is ​​binarized according to the segmentation threshold to generate a background mask matrix. Among them, when When, let the element value ;when When, let the element value Next, we introduce square structural elements. (i.e., a two-dimensional matrix with all matrix elements equal to 1). The size of this structuring element is designed to filter out minute noise at the boundaries while preserving the true boundary shape. To achieve adaptive filtering of the boundary shape, the size parameter of the structuring element is... Compared with the aforementioned local neighborhood window size Correlation, using formulas Dynamically determined. In this embodiment, corresponding to the aforementioned... The specific value after calculation is Using this structural element For the background mask matrix respectively Perform morphological dilation ( ) and corrosion ( The operation yields the morphological gradient matrix. The continuous pixel bands with a value of 1 in this matrix are the water-land boundary texture.

[0028] The land-water boundary region is determined based on the aforementioned land-water boundary texture. In specific implementation, the morphological gradient matrix is ​​extracted. The set of spatial coordinates corresponding to pixels with a median value of 1 is used to define the two-dimensional spatial range covered by this set in the original image as the land-water boundary region. The water body region is then determined based on this boundary region. Specifically, a mask matrix representing the non-water body shoreline is used. Based on this, interfering pixels in the water-land boundary region are removed using spatial truncation logic to generate a binary water mask matrix representing the water body region. The calculation formula is as follows: ; in, This is the function for maximizing the value. Within this water body mask matrix... In this context, the continuous region with an element value of 1 is the final determined water body region.

[0029] Obtain the drone heading angle recorded by the drone flight control system. With camera pan-tilt yaw angle When the drone is inspecting along the river channel and the camera pan-tilt unit is in a vertical downward view state, adjust the drone's heading angle. The physical direction is determined to be the actual spatial direction of the river channel. An initial color image is then established. The top left corner is the origin, and the horizontal direction to the right is... The positive direction of the axis is vertically downward. A two-dimensional image pixel coordinate system with the positive axis. When the UAV gimbal is in a vertical downward view state, the rotation angle of the river channel spatial direction in the two-dimensional image pixel coordinate system is calculated using a mapping formula. : ; in, This is the reference compensation angle between the image coordinate system and the gimbal physical coordinate system; in this embodiment, it is taken as [value missing]. The rotation angle is The positive direction of the axis is taken as The polar angle calculated from the baseline. (Based on the initial color image.) center pixel Based on this, establish a direction vector representing the spatial direction of the river channel. Along the direction vector According to the preset pixel step size Calculate the coordinates of each equally spaced division point: ; ; in, This is the floor function; For the initial color image The width; This is the step size scaling factor, used to adjust the region division step size according to the spatial size of different inspection images, so as to reduce the impact of changes in UAV acquisition distance and image resolution on the region division scale. Its preferred value range is... to In this embodiment, the specific value is... ; The integer step index has a range of values ​​set to 1. The formula for calculating boundary values ​​is: ; in, and The initial color images are respectively Width and height, This is the floor function. It passes through the coordinate point... And the normal vector and Equations of parallel lines As a region dividing line, the initial color image The region is divided into multiple contiguous candidate regions. A water mask matrix is ​​then used. Each candidate region is filtered, and water pixels with a mask value of 1 are retained within each candidate region to obtain multiple initial region images. Pixel connectivity detection is performed on each initial region image to remove isolated connected pixel regions and abnormal connected regions that are not water bodies, resulting in multiple water body region images.

[0030] Through the above processing, the water area was effectively extracted, the interference from the shoreline and vegetation was reduced, and a relatively stable data foundation was provided for subsequent adaptive image enhancement processing.

[0031] Furthermore, the process of removing isolated pixel connected blocks and non-water body abnormal pixel connected blocks from each initial region image includes: identifying pixel connected blocks in each initial region image based on the connectivity distance between adjacent pixels in each initial region image; obtaining the connectivity area, spatial location, and color features of the pixel connected blocks; removing pixel connected blocks with a connectivity area less than a preset area threshold as isolated pixel connected blocks; and removing pixel connected blocks whose spatial location is within the water-land boundary area, as well as pixel connected blocks whose color features deviate from the preset water body color range, as non-water body abnormal pixel connected blocks.

[0032] Specifically, after acquiring the initial region images, it is necessary to delineate independent pixel connected components based on the connectivity distance between adjacent pixels. In actual processing, the coordinates of any target pixel within the image are selected. As the center, delineate a A pixel-sized neighborhood window is used to lock the coordinates of eight adjacent pixels surrounding it in the horizontal, vertical, and diagonal directions. Considering physical boundary constraints, for a target pixel that is exactly at the edge of the image, only those neighboring pixels that actually exist are included in the calculation. In the RGB color space, if the color vector of this target pixel is defined as... The color vector of the corresponding adjacent pixel is denoted as Using the Euclidean distance model of three-dimensional vectors, the connectivity distance between these two elements in the color dimension can be calculated. The specific calculation expansion is as follows: ; Next, we proceed to the judgment stage, where a region connectivity threshold needs to be predefined. The intention behind setting this threshold is to accommodate slight color variations caused by normal water ripples while strictly preventing abrupt connections with the edges of foreign objects. A connectivity tolerance coefficient is specifically introduced into the formula. It is mainly used to compensate for differences in quantization depth (i.e., parameters) between different camera sensors. The differences in coefficients prevent numerical scaling errors during low-level calculations. The preferred value is 5% to 15%, and the specific value in this embodiment is [value missing]. Then, using a standard 8-bit image (corresponding to...) Taking the scenario as an example, The value is 25. When adjacent pixels belong to the same effective region after being cropped by the water mask matrix, and the connectivity distance satisfies... When the adjacent pixels are connected, it is determined that there is a valid connection relationship between them.

[0033] Next, a two-step connected component labeling algorithm is used to identify connected pixel components in the image. In actual processing, the first raster scan identifies adjacent pixels that satisfy the aforementioned valid connectivity relationship, assigns temporary labels to them, and records the equivalent pairs of labels. The second scan resolves these equivalent pairs, thereby reallocating globally unique labels. After these two scans, all sets of pixels with valid connectivity relationships are uniformly merged and labeled into independent connected pixel components (defined as connected pixel sets). Next, it is necessary to process each identified... Extracting multi-dimensional features. Specifically, this includes the following aspects: First, statistical analysis... The total number of pixels inside, which is used as the connected area. Secondly, to determine the physical location of this set, the arithmetic mean of the coordinates of all pixels within it is required, which will then be used as the geometric center coordinates. The specific mathematical expression is as follows: ; In addition, the color characteristics of this set in the RGB space need to be calculated. This involves obtaining the mean vector of the three channels. Taking the R channel as an example, its mean... The calculation formula is: ; Using the same calculation approach, the solution can be obtained similarly. and .

[0034] Then, based on the already calculated connected area Compared with the preset area threshold Conduct a comparison. Set up... The main purpose is to distinguish the main water area from the small color patches caused by local noise and small floating objects. Considering the optical resolution of conventional drone photography, this threshold is generally set as the initial color image size (width). high The percentage is between 0.1% and 0.5%. In this embodiment, the percentage is specifically set at 0.2%, and the calculated value is rounded down to the nearest integer. The final value. As long as it is discovered... This allows us to determine that the set consists of isolated pixel blocks that lack representativeness of water quality, and to remove them.

[0035] For the set of connected pixels that meet the area requirement and are retained Furthermore, a second verification must be performed on the spatial and color levels.

[0036] When performing spatial verification, a boundary buffer region needs to be constructed first. Since a fixed-pixel-width boundary buffer region is prone to causing spatial scale failures, a buffer band scaling factor needs to be introduced. And this coefficient is compared with the image width. After multiplying and rounding down, the pixel width of the boundary buffer region for adaptive image width can be obtained. The preferred range for this coefficient is between 0.5% and 1.5%; in this embodiment, it is specifically set to 1.0%. Then, the set of connected pixels is determined based on this width value. Geometric center coordinates Does it fall within the buffer zone extending from the land-water boundary into the water body? If it does, it indicates a high probability that the collection is mixed with shoreline background disturbances.

[0037] Color verification is used to determine whether the colors of connected pixel blocks deviate significantly from the preset water body color. During verification, the normalized color vector of the reference water body is first determined. Specifically, multiple historical images of the target water area were retrieved under conditions of normal water quality, natural lighting, and no obstructions. The three-channel color mean of the effective water area in each frame was calculated. The mean values ​​were then normalized using the maximum quantization extremum of the image acquisition device. The arithmetic mean of the normalization results corresponding to multiple historical images was then taken to obtain the normalized color vector of the reference water body. In this embodiment, Pick In actual processing, the image bit depth is taken into account. Using the mapping formula Convert it to a physical reference color vector For a standard 8-bit image, this physical reference color vector is taken as... Based on this, the color mean vector of the current connected pixel set is calculated. Color tolerance distance between the physical reference vector and the physical reference vector : ; Considering that different image bit depths will lead to differences in dimensions, a color tolerance scaling factor is introduced. The color tolerance threshold is dynamically determined according to the following formula: .

[0038] The preferred range is 10% to 20%, and in this embodiment, it is 15.6%. When the image bit depth is 8 bits, Approximately 40. This threshold, together with the aforementioned baseline color vector, defines the numerical range of normal water bodies to identify and filter out high-contrast abnormal blocks formed by factors such as dead leaves and fog. During the filtering process, if the set of connected pixels... Located in the buffer zone at the boundary between land and water, or its color tolerance distance If so, the connected pixel set is determined to be a non-water body abnormal pixel block and is removed.

[0039] Through the above-mentioned connected region filtering process, invalid pixels and abnormal non-water areas within the water body area are further cleaned up, reducing the impact of non-target areas on the calculation process of light attenuation parameters, near-water surface scattering parameters, and water surface reflection degradation parameters, and improving the accuracy of subsequent regional enhancement control parameter calculations.

[0040] Furthermore, referring to Figure 3 The specific process of extracting the illumination attenuation parameters, near-water surface scattering parameters, and water surface reflection degradation parameters of images of each water body region includes: acquiring the overall brightness distribution characteristics, local contrast attenuation characteristics, water surface transmission variation characteristics, high-brightness anomaly distribution characteristics, and directional reflection characteristics of images of each water body region; calculating the average brightness value based on the overall brightness distribution characteristics, and determining the illumination attenuation parameters based on the deviation of the average brightness value from the preset reference brightness value; determining the local contrast characterization quantity based on the local contrast attenuation characteristics, determining the transmittance spatial variation characterization quantity based on the water surface transmission variation characteristics, and weighting the local contrast characterization quantity and the transmittance spatial variation characterization quantity according to the preset scattering evaluation weight to determine the near-water surface scattering parameters; calculating the abnormal high-brightness pixel density based on the high-brightness anomaly distribution characteristics, determining the reflection direction intensity based on the directional reflection characteristics, and multiplying the abnormal high-brightness pixel density by the reflection direction intensity to determine the water surface reflection degradation parameters.

[0041] Specifically, for a space with dimensions of Gray quantization bit depth is The images of each of the aforementioned water body regions are represented as a three-channel image matrix in the RGB color space. The aforementioned weighted average calculation formula is used. For the three-channel image matrix Calculate pixel by pixel to obtain the corresponding region grayscale matrix. Because water surfaces often have specular reflection areas, some bright pixels may affect subsequent brightness statistics. Therefore, when obtaining the regional grayscale matrix... Then, based on the preset specular grayscale threshold Constructing a high-brightness mask matrix Highlighted areas are marked. Specifically, when... season ;when season Subsequently, the number of pixels with an element value of 1 in the mask matrix is ​​counted to obtain the total number of highlighted pixels. The preset high-gloss grayscale threshold is mentioned above. Used to distinguish between ordinary water surface brightness areas and bright reflection areas. Combined with image grayscale to quantize bit depth. The The preferred value range is set as follows For 8-bit quantized images ( In this embodiment, the value is 240.

[0042] Regarding the overall brightness distribution characteristics, after excluding the influence of bright areas, the regional grayscale matrix is ​​analyzed. Perform grayscale frequency statistics to generate a histogram vector representing the overall brightness distribution characteristics. Specifically, let the current image grayscale level be... Corresponding grayscale index Traverse the grayscale matrix of the region. China satisfies The effective pixels, the statistical grayscale value equals The number of pixels is used to obtain the histogram vector elements. Furthermore, normalization is performed based on the number of effective pixels to obtain the grayscale probability distribution. : ; in, This is a function for determining the maximum value, used to avoid calculation errors when the number of effective pixels is 0. Based on the stated grayscale probability distribution... Calculate the average brightness value of the region image. : ; Next, based on the average brightness value Compared with the preset reference brightness value The degree of deviation is used to determine the illumination attenuation parameter. : ; Wherein, the preset reference brightness value This represents the calibrated brightness value of a similar water body area under standard illumination conditions. To accommodate data with different gray quantization depths, its preferred value range is set to ( For 8-bit quantized images ( In this embodiment, the value is 140.

[0043] Next, the local contrast attenuation features in the water area image are quantified. Specifically, at the pixel level... Centered on, construct a structure with a size of Contrast Analysis Window Among them, window size Used to determine the local grayscale statistical range. To adapt to inspection images with different spatial resolutions, a value corresponding to the current water body area image size is introduced. Relevant proportional coefficients The window size is determined using the following formula: ; in, This is the floor function. In this embodiment, it takes... Traverse the grayscale matrix of the region. Set the traversal step size to 1 pixel, and expand outwards at the image boundary using pixel mirroring. Circle pixels. Perform statistical analysis on the pixels within a local window and calculate the average grayscale value within the window. : ; Further calculation of the local grayscale standard deviation yields the local standard deviation matrix representing the local contrast attenuation characteristics. : ; Calculate the local standard deviation matrix The global arithmetic mean is used to determine the local contrast characterization. : ; The water surface transmission variation characteristics were then quantified. Specifically, based on the three-channel image matrix... Use the same window size as the contrast analysis. Using edge expansion methods, the minimum intensity value of different color channels within each window is calculated to obtain the dark channel matrix. : ; in, Represents a three-channel image matrix Corresponding color channels The brightness value. To reduce the impact of bright reflection areas on the calculation of atmospheric light values, the aforementioned bright mask matrix is ​​used. Candidate regions are filtered. Specifically, the dark channel matrix is ​​filtered. The elements in the array are sorted in descending order of their numerical values, and elements that satisfy the condition are removed. The highlighted pixels are selected, and the top 0.1% of the remaining valid pixels are selected as the candidate set. Statistical candidate set Number of pixels .

[0044] Introduce a preset effective quantity threshold Perform prior applicability testing. Combined with the dynamic determination of image spatial dimensions, i.e. ,in As an effective proportionality coefficient, the value in this embodiment is [value to be filled in]. Determine the number of pixels. With effective quantity threshold Relationship: When When the dark channel prior is determined to be invalid in the current region, a degradation estimation failure state is triggered, and the transmittance matrix is ​​changed. All elements within are set to 1. When At that time, the average intensity of these candidate pixels in each channel is calculated to obtain the atmospheric light value vector. Among them, each channel component The calculation formula is: ; According to the atmospheric light value vector Calculate the pixel transmittance to obtain a transmittance matrix representing the characteristics of water surface transmittance variation. : ; in, To prevent division by zero anomalies, the extremely small constant is set, and in this embodiment, it is set to a value of [value missing]. . The preset lower limit of transmittance is used to limit the range of transmittance calculation. Its preferred value range is 0.05 to 0.15, and in this embodiment, the specific value is 0.1. A preset dehazing retention coefficient is used to adjust the degree of scattering correction; its preferred value range is 0.90 to 0.98, and in this embodiment, it is set to 0.95. Next, the aforementioned horizontal direction operator is used... with vertical direction operator For the transmittance matrix Perform two-dimensional discrete convolution to obtain the horizontal gradient of transmittance. and vertical gradient Calculate the square matrix of transmittance gradient magnitude based on the gradient. : ; Square matrix of transmittance gradient magnitude The arithmetic mean of all elements is used to obtain the spatial variation characteristic of transmittance. : ; Then, the local contrast characterization values ​​were respectively... Characterization of spatial variation of transmittance Normalization is performed to obtain the corresponding normalized local contrast characterization. Characterization of spatial variation of normalized transmittance The near-water surface scattering parameters are determined by linearly weighting the data according to preset scattering evaluation weights. : ; ; in, Current image bit depth The maximum theoretical grayscale range is 255 in an 8-bit image. and The preset scattering evaluation weights are used to adjust the influence of local contrast attenuation characteristics and spatial variation characteristics of transmittance on the calculation of scattering parameters. Their preferred value ranges are as follows: , In this embodiment, , .

[0045] Regarding the aforementioned abnormal distribution characteristics of high brightness, based on the total number of high brightness pixels obtained above... With the total number of pixels in the image The proportional relationship between them is used to calculate the density of abnormally bright pixels. : ; Finally, the quantization of directional reflection features is performed. Specifically, a model covering four directional angles is constructed. Gaber directional filter kernel The window size of the filter kernel is set to... A local coordinate system is established with the center of the window as the origin. For any coordinate point within the window... Its value range is The local coordinates are rotated based on the direction angle θ: ; The formulas for calculating the elements of the corresponding directional filter kernel matrix are as follows: ; in, and These are the local coordinate components after rotation; The standard deviation of the Gaussian envelope; The aspect ratio of the space; The wavelength is that of a sinusoidal carrier wave. This refers to the phase offset. Based on this, the filter kernel parameters are configured. The phase offset... Set to 0 to form an even-symmetric filter kernel. To adapt to inspection images with different spatial resolutions, the Gaussian envelope standard deviation is... With image width Establish a relationship: Its corresponding value range is 1.0 to 3.0 pixels, and in this embodiment it is taken as 1.0. Further setting the sinusoidal carrier wavelength with Gaussian envelope standard deviation Between In this embodiment Aspect Ratio This is used to adjust the response of the filter kernel to textures in different directions, and its preferred value range is 0.3 to 0.8. In this embodiment, it is taken as... The window size Based on Gaussian distribution The principle is determined, and the calculation formula is as follows: In this embodiment, the value is 13. Based on the above parameters, the corresponding Gabor directional filter kernels are generated for each direction. Then, respectively compared with the regional grayscale matrix Perform two-dimensional discrete convolution to obtain texture response matrices in each direction. : ; in, This represents the discrete convolution operation. It is derived from the aforementioned directional angles. Corresponding texture response matrix A set of multi-directional texture response matrices is constructed to represent the directional reflection features. Calculate the directional angles within the marked effective highlight area. The regional average response energy : ; Extract the average response energy of the region corresponding to the four orientation angles. The maximum value in the value determines the intensity of the reflection direction. : ; The abnormally high brightness pixel density Intensity in the direction of reflection Multiply to determine the water surface reflection degradation parameter. .

[0046] Through the above process, water image degradation features are extracted from regional brightness distribution, local contrast changes, water surface transmission changes, and high-brightness directional reflection, respectively, to obtain parameters that characterize the degree of influence of light attenuation, near-water surface scattering, and water surface reflection. These parameters can reflect the image changes of different water areas after being affected by insufficient light, scattering, and reflection, providing feature inputs for subsequent enhancement control parameter calculation and adaptive image enhancement processing.

[0047] Furthermore, referring to Figure 3 The process of polarity fusion of positive polarity illumination modulation features, negative polarity scattering modulation features, and negative polarity reflection modulation features to generate region enhancement control parameters includes: calculating the regional feature amplitudes of the positive polarity illumination modulation features, negative polarity scattering modulation features, and negative polarity reflection modulation features in the images of each water body region; normalizing each regional feature amplitude to obtain illumination weights, scattering weights, and reflection weights respectively; multiplying the positive polarity illumination modulation features by the illumination weights, the negative polarity scattering modulation features by the scattering weights, and the negative polarity reflection modulation features by the reflection weights, and algebraically superimposing the product results to obtain preliminary fusion parameters; and generating region enhancement control parameters based on the preliminary fusion parameters.

[0048] Specifically, obtain the light attenuation parameters calculated above. Near-water surface scattering parameters and water surface reflection degradation parameters Combining the regional grayscale matrix corresponding to each water body area image Transmittance matrix and high-brightness mask matrix Corresponding positive polarity illumination adjustment features, negative polarity scattering adjustment features, and negative polarity reflection adjustment features are constructed respectively. To achieve pixel-level spatial mapping, the above three features are explicitly represented as two-dimensional illumination adjustment feature matrices. Two-dimensional scattering modulated characteristic matrix and two-dimensional reflection modulation feature matrix For a space with dimensions of Quantization bit depth A water area image, traversing any coordinate point within it. Using depth extreme values The bound dimensionless formula calculates the corresponding eigenvalue matrix elements: ; ; ; The logic behind the above formula is as follows: For illumination compensation, pixels with more severe overall illumination loss and lower local grayscale are assigned a larger positive polarity adjustment reference amplitude; for scattering suppression, regions with more severe overall degradation and lower local transmittance are assigned a larger negative suppression feature amplitude; for reflection penalty, a high-brightness mask matrix is ​​used. The binary gating mechanism locks the overflowing pixels and amplifies the penalty amplitude according to their grayscale proportion. The two-dimensional illumination adjustment feature matrix is ​​traversed. Calculate the arithmetic mean of the absolute values ​​of all elements and use it as a scalar representing the magnitude of the positive polarity illumination modulation feature within the current image. The calculation formula is as follows: ; Using the same global arithmetic mean traversal method as described above, the two-dimensional scattering regulation feature matrix is ​​calculated respectively. With two-dimensional reflection modulation feature matrix The arithmetic mean of the absolute values ​​of all elements in the matrix is ​​used to obtain the corresponding negative polarity scattering characteristic amplitude scalar. With negative polarity reflection characteristic amplitude scalar Let the sum of the scalar magnitudes of the regional characteristics be... ,Right now The obtained characteristic amplitude scalars for each region are linearly normalized to determine the illumination weights. Scattering weights and reflection weight To ensure the smooth stability of the algorithm in the non-degenerate state, the following piecewise function is used for calculation: when hour: ; ; ; when Time (i.e., the non-degenerate ideal state where all characteristic amplitudes are zero): ; A pixel-level algebraic superposition with polarity labels is performed on the three two-dimensional spatial feature matrices to generate a two-dimensional preliminary parameter matrix representing the preliminary fusion parameters. Specifically regarding pixel coordinates : ; In this formula, the positive polarity illumination adjustment term is given a positive sign to implement positive enhancement, and the negative polarity scattering and reflection adjustment term is given a negative sign to implement reverse penalty, resulting in a preliminary parameter matrix. It can accurately reflect the net adjustment direction and magnitude after considering multiple physical degradation factors in space. Finally, regional enhancement control parameters are generated based on the preliminary fusion parameters.

[0049] By employing polarity weighting and algebraic fusion mapping of multidimensional degradation features, adaptive decoupling and antagonistic cancellation of complex physical degradation factors are achieved, providing a quantitative control benchmark refined to the pixel level for subsequent adaptive image reconstruction and feature extraction.

[0050] Furthermore, referring to Figure 4The specific process of generating region enhancement control parameters based on preliminary fusion parameters includes: determining the parameter change gradient based on the difference between the preliminary fusion parameters of spatially adjacent water body region images; when the parameter change gradient is greater than a preset gradient threshold, applying spatial continuity constraints to the preliminary fusion parameters to obtain region enhancement control parameters; when the parameter change gradient is not greater than the preset gradient threshold, using the preliminary fusion parameters as region enhancement control parameters.

[0051] Specifically, images of a first water body region and a second water body region adjacent to each other along the spatial direction of the river channel are acquired, and the corresponding preliminary fusion parameters are respectively represented as preliminary fusion parameter matrices. and The corresponding space dimensions are as follows: and For the seam between the first and second water body region images, preliminary fusion parameters are extracted from the first water body region image near the side boundary of the seam, and preliminary fusion parameters are also extracted from the corresponding side boundary of the seam in the second water body region image. The parameter differences at these corresponding locations are calculated, and the average is calculated along the seam direction to obtain the parameter change gradient, which represents the degree of parameter abruptness between adjacent water body region images. : ; in, Indicates the number of pixels in the direction of the region seam; This indicates the location of the first water body region image near the side boundary of the region seam. The initial fusion parameter values ​​corresponding to each pixel; This indicates the location of the second water body region image near the side boundary of the region seam. The initial fusion parameter values ​​corresponding to each pixel.

[0052] For images of other adjacent water body regions arranged continuously along the river channel spatial direction, the same method for calculating parameter differences at seam locations is used to determine the parameter variation gradient between corresponding regions. The calculated parameter variation gradient is then... With preset gradient threshold A comparison is performed. The preset gradient threshold... The setting is based on the upper limit of the parameter fault introduced by independent calculation between adjacent regions, and its preferred value range is 0.15 to 0.3. In this embodiment, the specific value is 0.2. When the parameter change gradient is greater than the preset gradient threshold, that is... This indicates a significant abrupt change in enhancement parameters at the seam between adjacent water body regions. Distance-weighted interpolation smoothing, serving as a spatial continuity constraint, is applied to the initial fusion parameters of the boundary region. Specifically, a width of [missing information] is established on the far right of the first water body region image. A first transition zone is established, and a second transition zone of equal width is established on the far left of the second water body region image. The width of the transition zone... The setting is based on balancing the smooth transition effect of the boundary with the degree of independent preservation of local features, and the calculation formula is: in, The preset transition band scaling factor is set to establish a smooth size that adapts to the image resolution and local region size. Its preferred value range is [value range missing]. to In this embodiment, the specific value is... The constant 3 is intended to ensure a minimum effective interpolation window at extremely low resolutions. Let the distance from a pixel within the transition zone to the horizontal pixel distance of the vertical seam be... ( ,in (Represents the boundary pixel closest to the seam). First, at the seam coordinates... At the joint, extract the rightmost parameter of the first water body region and the leftmost parameter of the second water body region to determine the common target parameter. : ; Let the ratio of the transition zone position be Using this positional scale, the initial fusion parameters in the adjacent transition regions of the first and second water body regions are smoothly interpolated towards the interior parameters to generate the corresponding region enhancement control parameter matrix. and : ; ; Using the above interpolation formula, when hour, The output values ​​on both sides of the joint are strictly equal to the common target parameter. This ensures that the regional boundaries are strictly equal in numerical value; when hour, The parameters are smoothly transitioned to the original initial fusion parameters for their respective regions. For pixels in non-boundary regions outside the transition zone, the original initial fusion parameters are kept unchanged. (when (time), and (when (Time). For other adjacent joints arranged along the spatial direction of the river channel, the same linear gradient interpolation method based on spatial distance is used for parameter continuity constraint. When the gradient of the parameter change is not greater than the preset gradient threshold, i.e. When this occurs, it indicates that the variation in enhancement control parameters between adjacent regions is within the allowable range, and the preliminary fusion parameters are completely retained as the region enhancement control parameters, i.e., let... and .

[0053] By detecting parameter changes in adjacent regions and processing spatial continuity constraints, the differences in enhancement parameters between adjacent water body images are adjusted to make parameter changes at the region boundaries smoother, reducing the impact of parameter abrupt changes during regional enhancement and providing continuous regional enhancement control parameters for subsequent image enhancement processing.

[0054] Furthermore, referring to Figure 4 The specific process of correcting the region enhancement control parameters based on water texture changes and edge gradient changes includes: extracting water texture features and edge gradient features from each water region image before and after enhancement; calculating the difference between the water texture features before and after enhancement to obtain the water texture change; calculating the difference between the edge gradient features before and after enhancement to obtain the edge gradient change; determining the texture protection component based on the deviation of the water texture change from a preset texture continuity threshold; determining the boundary perturbation degree based on the edge gradient change; and determining the boundary protection component based on the deviation of the boundary perturbation degree from preset first thresholds and preset second thresholds; wherein, the preset first threshold... If a threshold is less than a preset second threshold, and the boundary disturbance degree is not greater than a preset first threshold, the boundary protection component is the difference between the preset first threshold and the boundary disturbance degree; if the boundary disturbance degree is not less than the preset second threshold, the boundary protection component is the difference between the preset second threshold and the boundary disturbance degree; if the boundary disturbance degree is greater than the preset first threshold and less than the preset second threshold, the boundary protection component is zero; the texture protection component and the boundary protection component are weighted and summed according to preset correction weights to obtain the parameter correction amount; the parameter correction amount is algebraically added to the region enhancement control parameters to obtain the initial addition result; the corrected region enhancement control parameters are generated based on the initial addition result.

[0055] Specifically, obtain the grayscale matrix of the region corresponding to the aforementioned generated water area image. With two-dimensional enhanced control matrix Gray-level mapping is performed using pixel-by-pixel algebraic multiplication to generate a two-dimensional enhancement matrix representing the enhanced water body region image. : ; Water texture features and edge gradient features are extracted from the images of each water body region before and after enhancement. The same local entropy extraction logic and extreme value normalization method as in the aforementioned water body region segmentation step are used to obtain the grayscale matrix of the region before enhancement. Using the local entropy extrema as a common normalization benchmark, the gray-level matrix of the region is respectively... With the enhanced matrix Perform same-scale normalization to generate a two-dimensional matrix representing the pre-enhanced water texture features. A two-dimensional matrix representing the enhanced water texture features Similarly, using the same Sobel bidirectional partial derivative operator and normalization process as described above, the gray-level matrix of the region is processed respectively. With the enhanced matrix High-frequency edge extraction is performed to generate a two-dimensional matrix representing the gradient features of the enhanced front edge. A two-dimensional matrix representing the enhanced edge gradient features For any pixel coordinate point within the water body area. The absolute difference between the water texture features before and after enhancement is calculated to obtain the amount of water texture change, and a corresponding water texture change matrix is ​​generated according to spatial coordinates. : ; Using the same computational logic, the absolute difference between the edge gradient features before and after enhancement is calculated to obtain the edge gradient change, and the corresponding edge gradient change matrix is ​​generated. : ; The texture protection component is determined based on the deviation of the water texture variation amount from the preset texture continuity threshold. Specifically, the water texture variation matrix... The pixel values ​​in the data are compared with the preset texture continuity threshold. A pixel-by-pixel comparison is performed. When the change in water texture is no greater than the preset texture continuity threshold, it indicates that the enhancement process has not caused significant texture deformation, and the texture protection component at the corresponding position is set to zero. When the change in water texture is greater than the preset texture continuity threshold, it indicates that there is a risk of excessive texture stretching or local structural distortion at the corresponding position. A negative adjustment is generated to suppress the subsequent enhancement amplitude, and a two-dimensional matrix representing the texture protection component is generated. : ; Wherein, the preset texture continuity threshold The setting is based on the maximum reasonable deformation tolerance allowed in the normalized space for low-frequency textures such as natural water ripples, to avoid loss of physical features due to over-enhancement. Its preferred value range is 0.05 to 0.10, and in this embodiment, the specific value is 0.08.

[0056] The boundary perturbation degree is determined based on the change in edge gradient. Specifically, it is first determined using arbitrary pixel coordinates. Centered on, construct a space with dimensions of Local neighborhood pooling window To eliminate the influence of spatial resolution differences in inspection images on the local feature convergence scale, the pooling window size... It needs to be dynamically calculated based on the image size, using a formula. Confirmed, among which The preset pooling ratio coefficient has the following preferred value range: In this embodiment, the specific value is taken as Then, the gradient change matrix at the edges within the pooling window is... The arithmetic mean of the covered elements is calculated, and the result is used as the boundary perturbation degree of the center pixel, thereby generating the boundary perturbation degree matrix. : ; The boundary protection component is further determined based on the deviation of the boundary disturbance degree from the preset first threshold and the preset second threshold. Specifically, the preset first threshold is introduced. With the preset second threshold The preset first threshold The lower limit used to characterize the permissible edge variation is preferably in the range of 0.05 to 0.15, and in this embodiment, it is specifically set to 0.10; the preset second threshold The upper limit used to characterize the allowable edge variation is preferably in the range of 0.25 to 0.45, and in this embodiment, it is specifically set to 0.30. Both satisfy the following conditions. The constraint relationship is established. Based on the aforementioned dual thresholds, a piecewise function is used to calculate the coordinates of each point. The boundary protection component at the given location is used to generate a two-dimensional matrix representing that component. : ; Next, the texture protection component and the boundary protection component are weighted and summed according to preset correction weights to obtain the parameter correction amount. Specifically, the texture correction weights are set. With boundary correction weights Generate a correction matrix representing the parameter correction amount. : ; Preset correction weights are used to balance water surface texture preservation with edge oversharpening distortion constraints. Considering that false edges have a more significant impact on water quality image enhancement results, the weight allocation appropriately emphasizes boundary feature feedback. and The preferred value ranges are respectively , and satisfy In this embodiment, the specific value is... , The parameter correction amount is algebraically added to the area enhancement control parameters to obtain the initial addition result. The correction amount matrix is ​​then... With the enhanced control matrix Perform pixel-by-pixel addition to generate a two-dimensional initial matrix representing the initial addition result. : ; Finally, the corrected region enhancement control parameters are generated based on the initial summation results.

[0057] By extracting the texture and gradient differences before and after enhancement, and constructing a segmented feedback mechanism with a tolerance range, the above steps complete a feedback correction, which helps reduce the risk of losing subtle details of the water surface during enhancement, and to a certain extent negatively suppresses over-sharpening artifacts, providing a more stable data foundation for subsequent image processing and feature extraction.

[0058] Furthermore, referring to Figure 4 The specific process of generating corrected region enhancement control parameters based on the initial addition result includes: constraining the initial addition result according to a preset enhancement parameter constraint range to obtain constrained region enhancement control parameters; enhancing the water area image based on the constrained region enhancement control parameters to obtain a constrained enhanced image; extracting the real-time feature matrix corresponding to the constrained enhanced image and comparing the real-time feature matrix with the pre-stored reference water body feature matrix to obtain feature difference results; when the feature difference result is not greater than a preset difference upper limit, using the constrained region enhancement control parameters as corrected region enhancement control parameters; when the feature difference result is greater than a preset difference upper limit, using the region enhancement control parameters before algebraic addition as a benchmark, truncating the constrained region enhancement control parameters and using the truncated region enhancement control parameters as corrected region enhancement control parameters.

[0059] Specifically, the two-dimensional initial matrix obtained from the aforementioned calculation is used to represent the initial addition result. Based on the preset range of enhancement parameters, the two-dimensional initial matrix is... A threshold truncation is performed pixel by pixel to generate a two-dimensional constraint matrix that represents the constrained region enhancement control parameters. The calculation formula is: ; in, and These represent the lower and upper limits of the preset enhancement parameter constraint range, respectively. Their setting is based on the principle of forcing the multi-dimensional fused and superimposed values ​​to converge within the standard normalized dynamic range, preventing excessive compensation or penalty mapping from causing overflow or zeroing distortion of underlying pixel values. Their preferred value range is... In this embodiment, we take and Extract the aforementioned unprocessed grayscale measurement bit depth. Three-channel water area image matrix Using the two-dimensional constraint matrix A pixel-level multiplicative mapping is performed on it. To prevent overexposure overflow, a hardware extremum constraint at the current bit depth is applied simultaneously to generate a three-channel constraint-enhanced image matrix of the constrained enhanced image. Its various color channels The mapping formula is: ; After completing this mapping, in order to eliminate the interference of brightness enhancement on the optical color of water, the three-channel constrained enhancement image matrix needs to be modified. Perform a decoupled conversion from RGB to HSV color space. Specifically, for any pixel coordinates... First, divide the corresponding channel values ​​by Thus normalized to The interval is then used to extract the maximum channel component. Minimum channel component And calculate the channel range Next, the saturation of this pixel is calculated using the following conversion formula. With hue : ; ; Set the hue quantization level With saturation quantization series The hue quantization level The preferred value for balancing the resolution of the color spectrum with the computational dimension of the feature matrix is ​​between 16 and 72; in this embodiment, it is set to 36, meaning that each... It is divided into a hue range. The saturation quantization series... The distribution levels used for finely quantifying the color purity of water bodies are preferably in the range of 8 to 32, and are set to 10 in this embodiment. Next, each pixel in the constrained enhancement image is traversed. The hue index of each pixel in the two-dimensional histogram matrix is ​​determined by a floor-down mapping method. and saturation index The specific calculation formula is as follows: ; ; in, This is the floor function. This is the function for finding the minimum value.

[0060] After obtaining the hue index and saturation index corresponding to each pixel, the Kronecker impulse function is used. The number of pixels within each quantization interval is counted, and the statistical results are normalized to obtain the real-time feature matrix. The real-time feature matrix It is a two-dimensional normalized histogram matrix with spatial dimensions of... , of which line, number Column feature elements The calculation formula is: ; in, , Subsequently, the baseline water body feature matrix pre-stored in the airborne database is invoked. The baseline water body feature matrix is ​​calculated from multiple historical images of the same inspected water area under normal water quality and standard lighting conditions. Specifically, each historical image is processed using the same hue quantization level as described above. and saturation quantization series A two-dimensional color histogram is generated, and the mean of elements at corresponding positions in multiple historical histograms is calculated to obtain the baseline water body feature matrix. Furthermore, after acquiring a manually verified normal water quality calibration image, the corresponding generated two-dimensional histogram data is added to the historical sample set, and the average value of each element is recalculated to update the reference water body feature matrix. Next, the real-time feature matrix is ​​calculated. With reference water body characteristic matrix The spatial difference distance between them is used to obtain the feature difference result. The calculation formula is as follows: ; in, This indicates the degree of deviation between the color distribution corresponding to the current enhancement result and the color distribution of the normal water quality reference. A preset upper limit for difference can be further set. This is used to limit the range of color distribution changes allowed during the enhancement process, and its preferred value is between 0.15 and 0.25; in this embodiment, it is set to 0.20. When the feature difference result is not greater than the preset difference upper limit, i.e. When the current enhancement parameters do not cause a significant shift in the color distribution, the two-dimensional constraint matrix is... As a representation of the modified region enhancement control parameter, i.e. Conversely, if the feature difference result is greater than the preset difference upper limit, i.e. In this case, the enhanced control parameters of the region before algebraic addition are needed, i.e., the enhanced control matrix. Based on this, the constraint enhancement parameters are truncated to obtain the two-dimensional modified control matrix. : ; in, This is the proportional adjustment coefficient, used to limit the adjustment range of the constraint matrix relative to the original enhanced control matrix. Its preferred value range is 0.10 to 0.20, and in this embodiment, the value is 0.15.

[0061] Through the above processing, based on the difference between the color distribution of the images before and after enhancement and the normal water quality baseline distribution, the enhancement control parameters are restricted, thus establishing a constraint relationship between brightness adjustment and color preservation during the enhancement process. When the enhancement operation causes the color distribution change to exceed the preset range, the impact of color feature changes on the water quality discrimination results is reduced by decreasing the parameter adjustment amplitude.

[0062] This invention proposes an adaptive enhancement method for UAV imaging in low-light environments. The method first calculates enhancement parameters for different water body regions based on regional illumination attenuation, near-water surface scattering, and water surface reflection characteristics in the UAV-acquired images, and then adjusts the enhancement control parameters according to the differences in regional characteristics. During the enhancement parameter generation process, color distribution constraints are introduced to correct the enhanced image and limit color shifts caused during brightness enhancement. Simultaneously, by combining regional continuity constraints and parameter range limitations, the output image improves visibility under low-light conditions while preserving the original color characteristics of the water body regions, which can be used for subsequent water quality feature extraction and state analysis.

[0063] Example 2: The linear weighting and symmetrical distance calculation method used in Example 1 can enhance images of conventional low-light water areas. However, in cases of strong backlighting, concentrated highlights on the water surface, and significant changes in water flow, the linear constraint method may lead to uneven distribution of enhancement parameters or abrupt parameter changes at regional boundaries due to the large differences between different degradation factors.

[0064] Therefore, based on Embodiment 1, this embodiment further adjusts the enhancement parameter calculation process by introducing nonlinear weight normalization, structure-aware smoothing, and asymmetric feature distance calculation methods to adapt to the image enhancement needs in complex aquatic environments.

[0065] Furthermore, the process of normalizing the characteristic amplitudes of each region to obtain the illumination weight, scattering weight, and reflection weight includes: introducing a temperature adjustment factor and using a nonlinear normalization method based on exponential mapping to calculate the weights of the degradation characteristic amplitudes of each region.

[0066] Specifically, the characteristic amplitude scalar of the positive polarity illumination region obtained in Example 1 is extracted. negative polarity scattering characteristic amplitude scalar and negative polarity reflection characteristic amplitude scalar Based on the exponential mapping normalization formula, the illumination weights are calculated respectively. Scattering weights and reflection weight Among them, illumination weight The calculation formula is as follows: ; in, This is a temperature adjustment factor used to adjust the degree of weight variation among different degradation factors. The preferred value range is 0.5 to 1.5, and in this embodiment, the value is 1.2. Scattering weight With reflection weight The same method is used for calculation.

[0067] By using the above nonlinear normalization method, the weights corresponding to different degradation factors are dynamically adjusted according to the characteristic amplitude, thus avoiding a single high-amplitude degradation factor from directly occupying the main weight.

[0068] Furthermore, the specific process of applying spatial continuity constraints to the preliminary fusion parameters to obtain the regional enhancement control parameters includes: extracting local gray-level change features in the transition zone at the boundary of adjacent water bodies, and constructing a structural anisotropy factor based on the gray-level gradient direction; establishing an adaptive distance weight function by combining the distance relationship between the pixel and the boundary position of the region, and using the adaptive distance weight function to interpolate the preliminary fusion parameters of adjacent regions to obtain the regional enhancement control parameters.

[0069] Specifically, the parameter change gradient obtained in Example 1 is extracted. and the two-dimensional preliminary parameter matrix of two adjacent regions. and .when When it is determined that there is a significant parameter change between adjacent regions, with a horizontal adjacent width of... Structural constraints are applied within the transition zone at the boundary of the region. First, the grayscale matrix of the region generated in Example 1 is obtained. And the Sobel operator is used to calculate each pixel in the transition region. Corresponding horizontal grayscale gradient magnitude With vertical grayscale gradient magnitude Based on the horizontal and vertical gradient magnitudes, a structural anisotropy factor is constructed. This is used to describe the changes in the direction of water texture, and its calculation formula is as follows: ; in To prevent the introduction of a very small constant due to division by zero, this embodiment takes the value of Furthermore, an adaptive distance weighting function is constructed. : ; in, It is a sine function. The transition zone position ratio described in Example 1 This refers to the horizontal distance from the pixel within the transition zone to the seam, as described in Example 1. The adaptive distance weighting function is used... Instead of the original linear distance weights in Implementation Example 1, for and Cross-interpolation calculations are performed to obtain the regional enhancement control parameter matrix. and Its complete algebraic constraint formula is updated as follows: ; .

[0070] in, These are the common target parameters for the seams described in Example 1.

[0071] By using the above structural constraint method, the interpolation ratio at different locations is adjusted according to the direction of water texture, making the changes in enhancement parameters between adjacent regions more continuous.

[0072] Furthermore, the specific process of comparing the real-time feature matrix with the pre-stored reference water body feature matrix to obtain the feature difference results includes: performing dimensionality reduction statistics on the real-time feature matrix and the reference water body feature matrix respectively to obtain the hue direction distribution and saturation direction distribution; setting corresponding weight coefficients for different color components, and calculating the feature difference results based on the weighted distribution differences.

[0073] Specifically, the real-time feature matrix generated from the constraint-enhanced image in Example 1 is obtained. and the baseline water body feature matrix pre-stored in the airborne database. Since changes in hue typically correspond to changes in the color category of water, while changes in saturation mainly reflect the intensity of color, hue penalty coefficients are set separately for each. and saturation penalty coefficient This distinguishes the distribution variations across different dimensions. Among them, The preferred value range is 3.0 to 5.0, and the value in this embodiment is 4.0; The preferred value range is 0.5 to 1.0, and in this embodiment, the value is 1.0. First, the real-time feature matrix... Accumulation along the saturation direction yields the real-time hue edge distribution. : ; For real-time feature matrix Accumulation along the hue direction yields the real-time saturation edge distribution. : ; Similarly, for the baseline water body characteristic matrix Perform the same processing to obtain the corresponding baseline hue edge distribution. and baseline saturation edge distribution Based on the above edge distribution results, calculate the spatial difference distance. : ; The calculated Spatial difference distance in alternative embodiment one Following the threshold judgment method in Example 1, the two-dimensional constraint matrix is ​​selected or limited to obtain the final two-dimensional corrected control matrix. .

[0074] The above processing allows for comparison of the enhanced results with the color distribution under normal water quality conditions, and constraints are applied based on the degree of change in hue and saturation, thereby reducing color shifts during the adjustment of enhancement parameters.

[0075] Example 3: This embodiment is a supplementary explanation of the final image stitching output steps in the above embodiments.

[0076] Furthermore, the specific process of stitching together the images of each optimized water body region by region to output an optimized river water quality inspection image includes: obtaining non-water body pixels in the initial color image that are not labeled as water bodies and retaining their original color pixel values; using the global spatial coordinates of the initial color image as a reference, stitching together the images of each optimized water body region with the non-water body pixels that retain their original colors in situ to generate the final optimized river water quality inspection image.

[0077] Specifically, the initial color image is acquired. The complete set of pixel coordinates, and the water mask matrix from the aforementioned steps. The set of valid water body pixel coordinates calibrated. For the initial color image. The water body was not covered by the aforementioned mask matrix. The calibrated non-water body pixels (i.e., the background area corresponding to shoreline, vegetation, or non-water body anomalous pixel connected blocks) are completely preserved in the initial color image during the stitching process. The original color pixel values ​​in the image. (Based on the initial color image) Using the spatial physical coordinates as a global reference, the enhanced images of each optimized water body region are backfilled into the corresponding region in situ according to their corresponding spatial coordinates; and the non-water body pixels that retain their original pixel values ​​are fused and combined according to their original absolute coordinates. The final optimized river water quality inspection image is generated using the following spatial segmentation mapping formula. : in, For the global two-dimensional coordinates of the image, This represents the effective pixel value in the optimized water area image after polarity fusion and smoothing constraint processing.

[0078] The aforementioned spatial segmentation mapping mechanism effectively maintains the spatial topological continuity of the surrounding environment, such as shoreline and vegetation, by combining the optimized water area with the non-water background in situ under the absolute coordinate system, thus ensuring the coherence and integrity of the global scene semantics in the final output image.

[0079] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive enhancement method for UAV imaging in low-light environments, characterized in that, include: By collecting river water quality inspection images using drones, the spatial direction of the river channel and the texture of the water-land boundary are extracted. Based on the spatial direction of the river channel and the texture of the water-land boundary, the river water quality inspection images are divided into regions to obtain images of each water body region. Extract the illumination attenuation parameters, near-water surface scattering parameters, and water surface reflection degradation parameters from the images of each water body area; Positive polarity illumination modulation characteristics are generated based on illumination attenuation parameters, negative polarity scattering modulation characteristics are generated based on near-water surface scattering parameters, and negative polarity reflection modulation characteristics are generated based on water surface reflection degradation parameters. Polarity fusion is performed on the positive polarity illumination modulation characteristics, negative polarity scattering modulation characteristics, and negative polarity reflection modulation characteristics to generate regional enhancement control parameters; Based on the regional enhancement control parameters, the images of each water body region are enhanced, and the changes in water texture and edge gradient are obtained by comparing the images of each water body region before and after enhancement. The regional enhancement control parameters are corrected based on the changes in water texture and edge gradient. The images of each water body region are then re-enhanced based on the corrected regional enhancement control parameters to obtain optimized water body region images. The optimized water body region images are then stitched together by region to output optimized river water quality inspection images.

2. The adaptive enhancement method for UAV imaging in low-light environments according to claim 1, characterized in that, The process of dividing river water quality inspection images into regions based on the spatial direction of the river channel and the texture of the water-land boundary includes: extracting water texture features and edge gradient features from the river water quality inspection images; determining the texture of the water-land boundary based on the water texture features and edge gradient features; determining the water-land boundary region based on the water-land boundary region; determining the water body region range based on the water-land boundary region, and dividing the river water quality inspection images along the spatial direction of the river channel according to the water body region range to obtain each initial region image; removing isolated pixel connected blocks and non-water body abnormal pixel connected blocks from each initial region image to obtain each water body region image.

3. The adaptive enhancement method for UAV imaging in low-light environments according to claim 2, characterized in that, The process of removing isolated pixel connected blocks and non-water body abnormal pixel connected blocks from each initial region image includes: identifying pixel connected blocks in each initial region image based on the connection distance between adjacent pixels in each initial region image; obtaining the connection area, spatial location, and color features of the pixel connected blocks; removing pixel connected blocks with a connection area less than a preset area threshold as isolated pixel connected blocks; and removing pixel connected blocks whose spatial location is within the water-land boundary area, as well as pixel connected blocks whose color features deviate from the preset water body color range, as non-water body abnormal pixel connected blocks.

4. The adaptive enhancement method for UAV imaging in low-light environments according to claim 1, characterized in that, The specific process for extracting illumination attenuation parameters, near-water surface scattering parameters, and water surface reflection degradation parameters from images of each water body region includes: acquiring the overall brightness distribution characteristics, local contrast attenuation characteristics, water surface transmission variation characteristics, high-brightness anomaly distribution characteristics, and directional reflection characteristics of images of each water body region; calculating the average brightness value based on the overall brightness distribution characteristics, and determining the illumination attenuation parameters based on the deviation of the average brightness value from a preset reference brightness value; determining the local contrast characterization quantity based on the local contrast attenuation characteristics, determining the spatial variation characterization quantity of transmittance based on the water surface transmission variation characteristics, and weighting the local contrast characterization quantity and the spatial variation characterization quantity of transmittance according to a preset scattering evaluation weight to determine the near-water surface scattering parameters; calculating the density of abnormally bright pixels based on the high-brightness anomaly distribution characteristics, determining the reflection direction intensity based on the directional reflection characteristics, and multiplying the density of abnormally bright pixels by the reflection direction intensity to determine the water surface reflection degradation parameters.

5. The adaptive enhancement method for UAV imaging in low-light environments according to claim 1, characterized in that, The process of polarity fusion of positive polarity illumination modulation features, negative polarity scattering modulation features, and negative polarity reflection modulation features to generate region enhancement control parameters includes: calculating the regional feature amplitudes of the positive polarity illumination modulation features, negative polarity scattering modulation features, and negative polarity reflection modulation features in the images of each water body region; normalizing each regional feature amplitude to obtain illumination weights, scattering weights, and reflection weights respectively; multiplying the positive polarity illumination modulation features by the illumination weights, the negative polarity scattering modulation features by the scattering weights, and the negative polarity reflection modulation features by the reflection weights, and algebraically superimposing the product results to obtain preliminary fusion parameters; and generating region enhancement control parameters based on the preliminary fusion parameters.

6. The adaptive enhancement method for UAV imaging in low-light environments according to claim 5, characterized in that, The specific process of generating region enhancement control parameters based on preliminary fusion parameters includes: determining the parameter change gradient based on the difference between the preliminary fusion parameters of spatially adjacent water body region images; when the parameter change gradient is greater than a preset gradient threshold, applying spatial continuity constraints to the preliminary fusion parameters to obtain region enhancement control parameters; when the parameter change gradient is not greater than the preset gradient threshold, using the preliminary fusion parameters as region enhancement control parameters.

7. The adaptive enhancement method for UAV imaging in low-light environments according to claim 1, characterized in that, The specific process of correcting the region enhancement control parameters based on water texture changes and edge gradient changes includes: extracting water texture features and edge gradient features from each water region image before and after enhancement; calculating the difference between the water texture features before and after enhancement to obtain the water texture change; calculating the difference between the edge gradient features before and after enhancement to obtain the edge gradient change; determining the texture protection component based on the deviation of the water texture change from a preset texture continuity threshold; determining the boundary perturbation degree based on the edge gradient change; and determining the boundary protection component based on the deviation of the boundary perturbation degree from preset first thresholds and preset second thresholds; wherein, the preset first threshold... If the threshold is less than a preset second threshold, and the boundary disturbance degree is not greater than a preset first threshold, the boundary protection component is the difference between the preset first threshold and the boundary disturbance degree; if the boundary disturbance degree is not less than the preset second threshold, the boundary protection component is the difference between the preset second threshold and the boundary disturbance degree; if the boundary disturbance degree is greater than the preset first threshold and less than the preset second threshold, the boundary protection component is zero; the texture protection component and the boundary protection component are weighted and summed according to preset correction weights to obtain the parameter correction amount; the parameter correction amount is algebraically added to the region enhancement control parameters to obtain the initial addition result; the corrected region enhancement control parameters are generated based on the initial addition result.

8. The adaptive enhancement method for UAV imaging in low-light environments according to claim 7, characterized in that, The specific process of generating corrected region enhancement control parameters based on the initial addition result includes: constraining the initial addition result according to a preset enhancement parameter constraint range to obtain constrained region enhancement control parameters; enhancing the water area image based on the constrained region enhancement control parameters to obtain a constrained enhanced image; extracting the real-time feature matrix corresponding to the constrained enhanced image and comparing the real-time feature matrix with a pre-stored reference water body feature matrix to obtain feature difference results; when the feature difference results are not greater than a preset difference upper limit, using the constrained region enhancement control parameters as corrected region enhancement control parameters; when the feature difference results are greater than a preset difference upper limit, using the region enhancement control parameters before algebraic addition as a benchmark, truncating the constrained region enhancement control parameters and using the truncated region enhancement control parameters as corrected region enhancement control parameters.

Images