A low-light target detection method based on adaptive image enhancement

Abstract

The application discloses a low-light target detection method based on adaptive image enhancement, and belongs to the field of computer vision.The method comprises the following steps: 1, preprocessing an original low-light RGB image, and mapping the RGB image to an HSV color space; 2, clustering processing of a luminance V component is carried out by using a k-means algorithm; 3, joint feature extraction and weighted fusion are carried out on each V luminance subspace; 4, adaptive gamma parameters and enhancement weights of each subspace are solved, and region-by-region differential luminance correction is carried out on the three V luminance subspaces; 5, the optimized V luminance channel is reconstructed with an H color channel and an S saturation channel to obtain a complete HSV image, and the complete HSV image is converted into a final enhanced RGB image; and 6, the RGB image is input into a preset target detection network to complete detection.The application solves the problems of dark area detail loss and bright area overexposure distortion caused by global unified enhancement, and significantly improves the target feature expression capability and detection precision in a low-light scene.

Description

Technical Field

[0001] This invention belongs to the field of computer vision, specifically relating to a method for target detection in low-light images. Background Technology

[0002] Low-light target detection is a core research area in computer vision, widely used in scenarios such as nighttime security monitoring, autonomous driving, low-altitude detection, and outdoor visual inspection. Due to factors such as insufficient nighttime light, ambient stray light interference, and limitations in the light sensitivity of imaging equipment, images acquired in low-light scenarios generally suffer from problems such as low overall brightness, uneven distribution of light and dark areas, loss of detail in dark areas, overexposure in bright areas, high image noise density, and weakened target texture and edge features. This leads to traditional target detection models being prone to missed detections, false detections, large localization errors, and low classification accuracy, severely restricting the practical application of visual inspection systems in complex low-light scenarios.

[0003] Current mainstream low-light image enhancement and target detection technologies mostly employ global uniform brightness correction, fixed parameter filtering, and single feature extraction. These algorithms cannot adapt to the varying local illumination distribution characteristics of images. Global enhancement methods easily lead to insufficient brightening of dark areas and excessive enhancement of bright areas, while fixed parameter correction strategies struggle to accommodate the differences in brightness distribution across different low-light scenes. Simultaneously, traditional feature extraction methods can only capture shallow image features, failing to effectively extract faint target textures and edge information in low-light dark areas, resulting in low feature utilization and weak detail restoration capabilities. Furthermore, most enhancement algorithms tend to destroy the original image's color information during brightness optimization, causing color casts and artifacts, further reducing the accuracy and robustness of subsequent target detection.

[0004] In summary, existing low-light target detection technologies suffer from shortcomings such as poor illumination adaptability, insufficient detail reproduction, color distortion, and low detection accuracy, failing to meet the demands for high-precision and high-stability target detection in complex low-light scenarios. Therefore, this invention proposes a low-light target detection method based on adaptive image enhancement. This method addresses the deficiencies of existing technologies through a hierarchical brightness spatial clustering approach, multi-dimensional feature fusion, and regional adaptive brightness correction, effectively improving low-light image quality and target detection performance. Summary of the Invention

[0005] The purpose of this invention is to address the problems of existing low-light target detection methods, which do not perform fine-grained layering of image brightness space and mostly adopt global fixed parameter enhancement and single feature extraction methods, which cannot adapt to the local differential illumination distribution of images. These methods suffer from problems such as insufficient restoration of details in dark areas, easy overexposure in bright areas, easy color distortion, and insufficient extraction of weak target features. Therefore, this invention proposes a low-light target detection method based on adaptive image enhancement.

[0006] The specific implementation process of this invention is as follows: A low-light target detection method based on adaptive image enhancement, comprising the following steps: Step 1: Preprocess the original low-light RGB image to remove noise; map the preprocessed RGB image to the HSV color space to separate the independent hue (H) channel, saturation (S) channel, and brightness (V) channel. The specific steps are as follows: Let the original low-light RGB image be at any pixel position in the image Construct a 3×3 neighborhood pixel window centered on the target pixel, and extract the grayscale values ​​of all 9 pixels within the window. Sort the pixel values ​​in the set in ascending order and take the median value as the output value of the current pixel. The calculation formula is as follows: (1) in, For the median solution function, After median filtering Location pixel value; Median filtering results Adaptive Gaussian filtering is performed to remove Gaussian noise. Construct a 5×5 neighborhood window centered on the pixel, and calculate the mean and variance of the pixel grayscale values ​​within the window: (2) (3) in, The average pixel value of the local window. For local pixel variance, N =25 represents the total number of pixels in the window, and the Gaussian kernel parameters are dynamically mapped based on the local variance. (4) in, , representing the minimum filtering coefficient in the edge detail region; , representing the maximum filter coefficient in the flat noise region, based on adaptive Construct a two-dimensional Gaussian kernel function to perform image convolutional noise reduction: (5) (6) The filtered RGB image's pixel values ​​for each channel are linearly mapped from the original integer range of 0 to 255 to the floating-point range of [0,1], eliminating interference from differences in illumination amplitude. The formula is as follows: (7) For the preprocessed and normalized RGB image, the standard computer vision color space conversion formula is used to map it to the HSV color space pixel by pixel, and separate the three channels of hue (H), saturation (S), and brightness (V).

[0007] Step 2: Use the k-means algorithm to cluster the luminance V component, with k set to 3, corresponding to dark areas, medium bright areas and bright areas respectively; The specific steps are as follows: Construct a one-dimensional sample set of all pixel brightness values ​​in the V channel, where each sample is a pixel brightness value. The set of all samples is: (8) in, m Calculate the brightness value of a single sample based on the total number of pixels. With the k Cluster centers Euclidean distance: (9) in, For the first k Brightness-based cluster centers k =1, 2, 3, iterate through all pixels, and assign each brightness pixel to the cluster with the smallest distance. The classification formula is: (10).

[0008] Step 3) Perform joint feature extraction on each V brightness subspace obtained from clustering, and then perform weighted fusion; The specific steps are as follows: Assume the number of pixels in the brightness subspace to be processed is... The 3×3 and 5×5 dual-scale Sobel operators are used to calculate the horizontal and vertical gradients respectively, and the gradient magnitude features are synthesized to adapt to edge structures of different thicknesses. (11) (12) in, For convolution operations, and These represent the Sobel horizontal and vertical gradient operators, respectively. The formula for the single-pixel gradient magnitude feature is as follows: (13) By fusing the gradient magnitudes at both scales, the final multi-scale gradient features are obtained: (14) in, The gradient magnitude is on a 3×3 scale. The gradient magnitude is on a 5×5 scale. These are the weighting coefficients; An 8-neighborhood uniform LBP operator is used to extract subtle texture features in low-light dark areas, focusing on the center pixel. Traverse 8 neighboring pixels Binary encoding rules: (15) LBP texture encoding value calculation formula: (16) To adapt to weak texture features in low-light conditions, the original LBP features are normalized to obtain standardized texture features: (17) in, , These are the maximum and minimum values ​​of the LBP features in the current luminance subspace, respectively, and the texture features are normalized to the [0,1] interval; For each luminance subspace, gradient structure features and texture detail features are adaptively weighted and fused, with weights assigned according to the region's luminance characteristics: texture weights are increased in extremely dark areas, and gradient weights are increased in bright areas. The joint feature formula is as follows: (18) in, k =1, 2, 3 correspond to three brightness subspaces; This represents the subspace texture weight.

[0009] Step 4) Solve for the adaptive gamma parameters and enhancement weights of each subspace, and perform region-by-region differential brightness correction on the three V brightness subspaces to complete the overall enhancement of the V channel; The specific steps are as follows: Regarding the first k Subspace Calculate the mean brightness of the subspace to characterize the overall illumination level of the region: (19) in, For subspace Total number of pixels, set the global optimal target brightness benchmark value Calculate the brightness deviation: (20) when When this occurs, it indicates that the area is too dark and needs to be brightened. When the area is too bright, it needs to be suppressed to reduce overexposure; Using the subspace joint features extracted in step 3) Construct a regional feature sharpness factor to characterize the richness of detail and avoid excessive enhancement and amplification of noise in low-light areas: (twenty one) in, Representing the k Subspace feature sharpness. Combining brightness bias and feature sharpness, a dual-constraint adaptive gamma parameter model is constructed: (twenty two) in, The adjustment coefficient is used to control the enhancement amplitude and prevent excessive distortion. For the first k Adaptive gamma parameters for a luminance subspace; To ensure a smooth transition at the boundary after zonal enhancement, a regional enhancement weight coefficient is introduced, weakening the enhancement weight for high-brightness areas and strengthening the repair weight for low-brightness areas. (twenty three) in, Weights are increased for subspaces; the lower the brightness, the greater the weight. For those belonging to the first k Subspace For each pixel, non-linear brightness mapping is performed using corresponding adaptive gamma parameters. The basic enhancement formula is: (twenty four) in, The brightness pixel value after gamma correction for a single zone; By combining formulas (23) and (24), the enhancement results of each partition are weighted and fused to balance the enhancement intensity of different regions: (25) in, This is the initial enhanced brightness image after multi-region weighted fusion. This is a region indicator function used to determine the subspace to which a pixel belongs: (26) To address the issues of abrupt brightness changes and edge artifacts that easily occur at the boundaries of cluster partitions, a neighborhood mean smoothing constraint is used to perform local smoothing optimization on the initially enhanced image: (27) Among them, the total number of pixels in the window N =25. To ensure that the enhanced pixels always meet the normalization interval specification and avoid numerical overflow, a range truncation constraint is applied to the output result: (28).

[0010] Step 5) Reconstruct the optimized V luminance channel with the original, unmodified H hue channel and S saturation channel to obtain the enhanced complete HSV image. Then, convert the reconstructed HSV image into the final enhanced RGB image using inverse color space transformation. The specific steps are as follows: Constructing the enhanced HSV image tensor Channel combination rules: (29) Will Restore the RGB three-channel components pixel by pixel: (30) in, This represents the converted RGB image. This represents the transformation function.

[0011] Step 6) Input the final enhanced RGB image into the preset target detection network to complete the detection of various targets in low-light scenes; The specific steps are as follows: Input the enhanced image obtained in step 5) into the YOLO object detection network, and output the predicted scores for each category through forward inference: (31) in, To determine the optimal classification category for the target in the final determination, the category with the highest probability is selected as the classification result for this low-light target detection. This represents the target's original predicted score output by the YOLO network.

[0012] The beneficial effects of this invention are: Existing low-light target detection methods lack fine-grained zoning of image brightness, generally employing global fixed-parameter enhancement and single-feature extraction. This fails to adapt to the local brightness differences in images, leading to severe loss of detail in dark areas, overexposure distortion in bright areas, and poor representation of weak target features. Ultimately, this results in low target detection accuracy and high false negative / missing rates in low-light scenarios. To address this, this invention converts low-light RGB images to the HSV color space, optimizing only the luminance (V) channel while preserving original hue and saturation information. Then, K-means clustering is used to adaptively layer the luminance channels, dividing them into multiple luminance subspaces. Next, multi-scale gradient-texture joint feature extraction is performed on each luminance subspace, fusing multi-dimensional feature information and constructing a feature feedback mechanism to adaptively solve for the optimal gamma enhancement parameters for each subspace, achieving differentiated luminance correction across regions. Finally, the enhanced luminance channels are reconstructed with the original H and S channels and converted back to RGB images, which are then input into a YOLO network for target detection.

[0013] This invention takes into account the brightness differences and weak feature distribution characteristics of different regions in low-light images, and adopts a brightness clustering and hierarchical and feature feedback adaptive enhancement strategy. It can effectively solve the problems of missing details in dark areas and overexposure distortion in bright areas caused by global uniform enhancement, and significantly improve the target feature expression ability and detection accuracy in low-light scenes. Attached Figure Description

[0014] Figure 1 This is a flowchart of the algorithm of the present invention; Figure 2 It is an unprocessed low-light image; Figure 3 These are the target detection results for unprocessed low-light images; Figure 4 It is a processed low-light image; Figure 5 This is the target detection result of the processed low-light image. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0016] A low-light target detection method based on adaptive image enhancement, the process is as follows: Figure 1 As shown, the steps are as follows: Step 1: Preprocess the original low-light RGB image to remove noise; map the preprocessed RGB image to the HSV color space to separate the independent hue (H) channel, saturation (S) channel, and brightness (V) channel. The specific steps are as follows: Let the original low-light RGB image be at any pixel position in the image Construct a 3×3 neighborhood pixel window centered on the target pixel, and extract the grayscale values ​​of all 9 pixels within the window. Sort the pixel values ​​in the set in ascending order and take the median value as the output value of the current pixel. The calculation formula is as follows: (1) in, For the median solution function, After median filtering Location pixel value; Median filtering results Adaptive Gaussian filtering is performed to remove Gaussian noise. Construct a 5×5 neighborhood window centered on the pixel, and calculate the mean and variance of the pixel grayscale values ​​within the window: (2) (3) in, The average pixel value of the local window. For local pixel variance, N =25 represents the total number of pixels in the window, and the Gaussian kernel parameters are dynamically mapped based on the local variance. (4) in, , representing the minimum filtering coefficient in the edge detail region; , representing the maximum filter coefficient in the flat noise region, based on adaptive Construct a two-dimensional Gaussian kernel function to perform image convolutional noise reduction: (5) (6) The filtered RGB image's pixel values ​​for each channel are linearly mapped from the original integer range of 0 to 255 to the floating-point range of [0,1], eliminating interference from differences in illumination amplitude. The formula is as follows: (7) For the preprocessed and normalized RGB image, the standard computer vision color space conversion formula is used to map it to the HSV color space pixel by pixel, and separate the three channels of hue (H), saturation (S), and brightness (V).

[0017] Step 2: Use the k-means algorithm to cluster the luminance V component, with k set to 3, corresponding to dark areas, medium bright areas and bright areas respectively; The specific steps are as follows: Construct a one-dimensional sample set of all pixel brightness values ​​in the V channel, where each sample is a pixel brightness value. The set of all samples is: (8) in, m Calculate the brightness value of a single sample based on the total number of pixels. With the k Cluster centers Euclidean distance: (9) in, For the first k Brightness-based cluster centers k =1, 2, 3, iterate through all pixels, and assign each brightness pixel to the cluster with the smallest distance. The classification formula is: (10).

[0018] Step 3) Perform joint feature extraction on each V brightness subspace obtained from clustering, and then perform weighted fusion; The specific steps are as follows: Assume the number of pixels in the brightness subspace to be processed is... The 3×3 and 5×5 dual-scale Sobel operators are used to calculate the horizontal and vertical gradients respectively, and the gradient magnitude features are synthesized to adapt to edge structures of different thicknesses. (11) (12) in, For convolution operations, and These represent the Sobel horizontal and vertical gradient operators, respectively. The formula for the single-pixel gradient magnitude feature is as follows: (13) By fusing the gradient magnitudes at both scales, the final multi-scale gradient features are obtained: (14) in, The gradient magnitude is on a 3×3 scale. The gradient magnitude is on a 5×5 scale. These are the weighting coefficients; An 8-neighborhood uniform LBP operator is used to extract subtle texture features in low-light dark areas, focusing on the center pixel. Traverse 8 neighboring pixels Binary encoding rules: (15) LBP texture encoding value calculation formula: (16) To adapt to weak texture features in low-light conditions, the original LBP features are normalized to obtain standardized texture features: (17) in, , These are the maximum and minimum values ​​of the LBP features in the current luminance subspace, respectively, and the texture features are normalized to the [0,1] interval; For each luminance subspace, gradient structure features and texture detail features are adaptively weighted and fused, with weights assigned according to the region's luminance characteristics: texture weights are increased in extremely dark areas, and gradient weights are increased in bright areas. The joint feature formula is as follows: (18) in, k =1, 2, 3 correspond to three brightness subspaces; This represents the subspace texture weight.

[0019] Step 4) Solve for the adaptive gamma parameters and enhancement weights of each subspace, and perform region-by-region differential brightness correction on the three V brightness subspaces to complete the overall enhancement of the V channel; The specific steps are as follows: Regarding the first k Subspace Calculate the mean brightness of the subspace to characterize the overall illumination level of the region: (19) in, For subspace Total number of pixels, set the global optimal target brightness benchmark value Calculate the brightness deviation: (20) when When this occurs, it indicates that the area is too dark and needs to be brightened. When the area is too bright, it needs to be suppressed to reduce overexposure; Using the subspace joint features extracted in step 3) Construct a regional feature sharpness factor to characterize the richness of detail and avoid excessive enhancement and amplification of noise in low-light areas: (twenty one) in, Representing the k Subspace feature sharpness. Combining brightness bias and feature sharpness, a dual-constraint adaptive gamma parameter model is constructed: (twenty two) in, The adjustment coefficient is used to control the enhancement amplitude and prevent excessive distortion. For the first k Adaptive gamma parameters for a luminance subspace; To ensure a smooth transition at the boundary after zonal enhancement, a regional enhancement weight coefficient is introduced, weakening the enhancement weight for high-brightness areas and strengthening the repair weight for low-brightness areas. (twenty three) in, Weights are increased for subspaces; the lower the brightness, the greater the weight. For those belonging to the first k Subspace For each pixel, non-linear brightness mapping is performed using corresponding adaptive gamma parameters. The basic enhancement formula is: (twenty four) in, The brightness pixel value after gamma correction for a single zone; By combining formulas (23) and (24), the enhancement results of each partition are weighted and fused to balance the enhancement intensity of different regions: (25) in, This is the initial enhanced brightness image after multi-region weighted fusion. This is a region indicator function used to determine the subspace to which a pixel belongs: (26) To address the issues of abrupt brightness changes and edge artifacts that easily occur at the boundaries of cluster partitions, a neighborhood mean smoothing constraint is used to perform local smoothing optimization on the initially enhanced image: (27) Among them, the total number of pixels in the window N =25. To ensure that the enhanced pixels always meet the normalization interval specification and avoid numerical overflow, a range truncation constraint is applied to the output result: (28).

[0020] Step 5) Reconstruct the optimized V luminance channel with the original, unmodified H hue channel and S saturation channel to obtain the enhanced complete HSV image. Then, convert the reconstructed HSV image into the final enhanced RGB image using inverse color space transformation. The specific steps are as follows: Constructing the enhanced HSV image tensor Channel combination rules: (29) Will Restore the RGB three-channel components pixel by pixel: (30) in, This represents the converted RGB image. This represents the transformation function.

[0021] Step 6) Input the final enhanced RGB image into the preset target detection network to complete the detection of various targets in low-light scenes; The specific steps are as follows: Input the enhanced image obtained in step 5) into the YOLO object detection network, and output the predicted scores for each category through forward inference: (31) in, To determine the optimal classification category for the target in the final determination, the category with the highest probability is selected as the classification result for this low-light target detection. This represents the target's original predicted score output by the YOLO network.

[0022] Example 1: As shown in Figure 2, this is the original low-light indoor furniture image without the adaptive enhancement processing of this invention. The bedside lamp is a bright area, while the area under the bed, the corners of the cabinets, and the textured areas of the cushions are dark areas. The overall brightness distribution of the original image is extremely uneven: the textures of the cushions and bedside tables in the dark areas are masked by noise, resulting in low detail recognition; the brightness of the lampshade area is concentrated, posing a risk of local overexposure. The image as a whole has not undergone regional brightness optimization, resulting in weak target features in dark areas and compressed pixel information in bright areas. Directly inputting this original low-light image into the YOLO object detection network yields the detection results shown in Figure 3. The blue boxes in the figure represent the target detection boxes output by the network, corresponding to several target categories: vase, bedside tables on both sides, leaf-shaped cushions on the left and right, lamp, and carpet at the foot of the bed. It can be seen that if target detection is directly performed on the unprocessed low-light image, there will be significant missed detections: the texture of the black and white patterned cushion in the center of the image is dim and its features are not prominent, and the network does not generate corresponding detection boxes, resulting in missed target detections.

[0023] Figure 4 shows a low-light image optimized using the adaptive image enhancement low-light processing method of this invention. The algorithm first performs median filtering and adaptive Gaussian joint noise reduction on the original image, and converts it to HSV space to separate the brightness V channel. Using K-means, the V channel pixels are divided into three subspaces: dark, medium-bright, and bright. Joint features are extracted by dual-scale Sobel gradient and 8-neighborhood LBP texture. Feature weights are adaptively assigned according to the brightness characteristics of each region. The adaptive gamma correction parameters of the partition are solved by combining brightness deviation and sharpness factor. The dark areas are brightened differently, the overexposure of the bright areas of the desk lamp is suppressed, the partition boundaries are smoothed to eliminate artifacts, and finally the original H and S channels are fused to restore unbiased colors. The final image has balanced brightness and darkness: the texture of the pillow in the dark area is clearly distinguishable, the details of the bright areas of the desk lamp are well preserved, the whole image has no obvious color cast, noise is effectively suppressed, and the edge and texture features of various furniture objects are fully restored.

[0024] The enhanced RGB image in Figure 4 was input into the same YOLO object detection network, and the detection results are shown in Figure 5, where the blue boxes represent the object detection boxes. Comparing the detection results with the original image clearly demonstrates the improvement effect of this invention: a new, accurate detection box for the central black-and-white patterned pillow was added, eliminating missed detections; vases, bedside tables, pillows, lamps, and carpets were all completely detected, and all detection boxes showed a higher degree of fit with the true contours of the objects, significantly reducing positioning errors; the classification confidence of small objects in dark areas and targets with weak textures was significantly improved, with no invalid false detection boxes. This fully verifies that this invention can effectively improve the quality of low-light images and significantly enhance the completeness and accuracy of object detection in low-light scenes.

Claims

1. A low-light target detection method based on adaptive image enhancement, characterized in that, The steps are as follows: Step 1: Preprocess the original low-light RGB image to remove noise; map the preprocessed RGB image to the HSV color space to separate the independent hue (H) channel, saturation (S) channel, and brightness (V) channel. Step 2: Use the k-means algorithm to cluster the luminance V component, with k set to 3, corresponding to dark areas, medium bright areas and bright areas respectively; Step 3) Perform joint feature extraction on each V brightness subspace obtained from clustering, and then perform weighted fusion; Step 4) Solve for the adaptive gamma parameters and enhancement weights of each subspace, and perform region-by-region differential brightness correction on the three V brightness subspaces to complete the overall enhancement of the V channel; Step 5) Reconstruct the optimized V luminance channel with the original, unmodified H hue channel and S saturation channel to obtain the enhanced complete HSV image. Then, convert the reconstructed HSV image into the final enhanced RGB image using inverse color space transformation. Step 6) Input the final enhanced RGB image into the preset target detection network to complete the detection of various targets in low-light scenes.

2. The low-light target detection method based on adaptive image enhancement according to claim 1, characterized in that: In step 1), the specific steps are as follows: Let the original low-light RGB image be at any pixel position in the image Construct a 3×3 neighborhood pixel window centered on the target pixel, and extract the grayscale values ​​of all 9 pixels within the window. Sort the pixel values ​​in the set in ascending order and take the median value as the output value of the current pixel. The calculation formula is as follows: (1) in, For the median solution function, After median filtering Location pixel value; Median filtering results Adaptive Gaussian filtering is performed to remove Gaussian noise. Construct a 5×5 neighborhood window centered on the pixel, and calculate the mean and variance of the pixel grayscale values ​​within the window: (2) (3) in, The average pixel value of the local window. For local pixel variance, N =25 represents the total number of pixels in the window, and the Gaussian kernel parameters are dynamically mapped based on the local variance. (4) in, , representing the minimum filtering coefficient in the edge detail region; , representing the maximum filter coefficient in the flat noise region, based on adaptive Construct a two-dimensional Gaussian kernel function to perform image convolutional noise reduction: (5) (6) The filtered RGB image's pixel values ​​for each channel are linearly mapped from the original integer range of 0 to 255 to the floating-point range of [0,1], eliminating interference from differences in illumination amplitude. The formula is as follows: (7) For the preprocessed and normalized RGB image, the standard computer vision color space conversion formula is used to map it to the HSV color space pixel by pixel, and separate the three channels of hue (H), saturation (S), and brightness (V).

3. The low-light target detection method based on adaptive image enhancement according to claim 1, characterized in that: In step 2), the specific steps are as follows: Construct a one-dimensional sample set of all pixel brightness values ​​in the V channel, where each sample is a pixel brightness value. The set of all samples is: (8) in, m Calculate the brightness value of a single sample based on the total number of pixels. With the k Cluster centers Euclidean distance: (9) in, For the first k Brightness-based cluster centers k =1, 2, 3, iterate through all pixels, and assign each brightness pixel to the cluster with the smallest distance. The classification formula is: (10)。 4. The low-light target detection method based on adaptive image enhancement according to claim 1, characterized in that: In step 3), the specific steps are as follows: Assume the number of pixels in the brightness subspace to be processed is... The 3×3 and 5×5 dual-scale Sobel operators are used to calculate the horizontal and vertical gradients respectively, and the gradient magnitude features are synthesized to adapt to edge structures of different thicknesses. (11) (12) in, For convolution operations, and These represent the Sobel horizontal and vertical gradient operators, respectively. The formula for the single-pixel gradient magnitude feature is as follows: (13) By fusing the gradient magnitudes at both scales, the final multi-scale gradient features are obtained: (14) in, The gradient magnitude is on a 3×3 scale. The gradient magnitude is on a 5×5 scale. These are the weighting coefficients; An 8-neighborhood uniform LBP operator is used to extract subtle texture features in low-light dark areas, focusing on the center pixel. Traverse 8 neighboring pixels Binary encoding rules: (15) LBP texture encoding value calculation formula: (16) To adapt to weak texture features in low-light conditions, the original LBP features are normalized to obtain standardized texture features: (17) in, , These are the maximum and minimum values ​​of the LBP features in the current luminance subspace, respectively, and the texture features are normalized to the [0,1] interval; For each luminance subspace, gradient structure features and texture detail features are adaptively weighted and fused, with weights assigned according to the region's luminance characteristics: texture weights are increased in extremely dark areas, and gradient weights are increased in bright areas. The joint feature formula is as follows: (18) in, k =1, 2, 3 correspond to three brightness subspaces; This represents the subspace texture weight.

5. The low-light target detection method based on adaptive image enhancement according to claim 1, characterized in that: In step 4), the specific steps are as follows: Regarding the first k Subspace Calculate the mean brightness of the subspace to characterize the overall illumination level of the region: (19) in, For subspace Total number of pixels, set the global optimal target brightness benchmark value Calculate the brightness deviation: (20) when When this occurs, it indicates that the area is too dark and needs to be brightened. When the area is too bright, it needs to be suppressed to reduce overexposure; Using the subspace joint features extracted in step 3) Construct a regional feature sharpness factor to characterize the richness of detail and avoid excessive enhancement and amplification of noise in low-light areas: (21) in, Representing the k Subspace feature sharpness. Combining brightness bias and feature sharpness, a dual-constraint adaptive gamma parameter model is constructed: (22) in, The adjustment coefficient is used to control the enhancement amplitude and prevent excessive distortion. For the first k Adaptive gamma parameters for a luminance subspace; To ensure a smooth transition at the boundary after zonal enhancement, a regional enhancement weight coefficient is introduced, weakening the enhancement weight for high-brightness areas and strengthening the repair weight for low-brightness areas. (23) in, Weights are increased for subspaces; the lower the brightness, the greater the weight. For those belonging to the first k Subspace For each pixel, non-linear brightness mapping is performed using corresponding adaptive gamma parameters. The basic enhancement formula is: (24) in, The brightness pixel value after gamma correction for a single zone; By combining formulas (23) and (24), the enhancement results of each partition are weighted and fused to balance the enhancement intensity of different regions: (25) in, This is the initial enhanced brightness image after multi-region weighted fusion. This is a region indicator function used to determine the subspace to which a pixel belongs: (26) To address the issues of abrupt brightness changes and edge artifacts that easily occur at the boundaries of cluster partitions, a neighborhood mean smoothing constraint is used to perform local smoothing optimization on the initially enhanced image: (27) Among them, the total number of pixels in the window N =25. To ensure that the enhanced pixels always meet the normalization interval specification and avoid numerical overflow, a range truncation constraint is applied to the output result: (28)。 6. The low-light target detection method based on adaptive image enhancement according to claim 1, characterized in that: In step 5), the specific steps are as follows: Constructing the enhanced HSV image tensor Channel combination rules: (29) Will Restore the RGB three-channel components pixel by pixel: (30) in, This represents the converted RGB image. This represents the transformation function.

7. The low-light target detection method based on adaptive image enhancement according to claim 6, characterized in that: In step 6), the specific steps are as follows: Input the enhanced image obtained in step 5) into the YOLO object detection network, and output the predicted scores for each category through forward inference: (31) in, To determine the optimal classification category for the target in the final determination, the category with the highest probability is selected as the classification result for this low-light target detection. This represents the target's original predicted score output by the YOLO network.

Images