Low-illumination dynamic vision SLAM (Simultaneous Localization and Mapping) method and system fusing robust features

Abstract

The invention discloses a robust feature fused low-illumination dynamic vision SLAM method and system. The method comprises the steps of performing fast non-local mean denoising on an input low-illumination original RGB image; then adaptively and sequentially applying a Zero-DCE network to carry out global brightness enhancement and CLAHE (contrast limited adaptive histogram equalization) to carry out local detail enhancement according to the average brightness of the image; on the basis, key points are extracted by adopting a self-adaptive FAST corner detection algorithm for dynamically adjusting a threshold value based on an image gray average value; identifying and removing feature points in dynamic areas such as pedestrians and vehicles by using a YOLOv8-Seg semantic segmentation model; and finally, the original ORB descriptor is upgraded into an enhanced descriptor with high robustness through a FeatureBooster deep learning module. According to the method, the stability and accuracy of feature extraction and matching under low illumination, motion blur and dynamic interference can be remarkably improved, and then the pose estimation precision and map construction integrity of an SLAM system are improved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and robotics, and in particular to a low-light dynamic visual SLAM method and system that integrates robust features. Background Technology

[0002] Visual SLAM is one of the core technologies for realizing applications such as autonomous navigation of mobile robots, path planning of drones, and augmented reality (AR). Currently, mainstream visual SLAM systems such as ORB-SLAM3 and VINS-Mono perform excellently in ideal environments (i.e., well-lit and static scenes). However, these systems are generally based on two key assumptions:

[0003] 1. The ambient lighting is good and constant;

[0004] 2. The vast majority of objects in the scene are static.

[0005] In real-world applications, these assumptions are often broken. For example, in low-light conditions such as at night, in tunnels, in backlight, or with high dynamic range (HDR), images captured by traditional cameras often have low contrast and high noise. This causes a sharp decline in the performance of feature extractors based on local grayscale changes (such as SIFT and ORB based on FAST corner detection) or even their failure, leading to problems such as tracking loss and relocalization failure. At the same time, in densely populated areas such as shopping malls and streets, a large number of dynamic objects (such as pedestrians and vehicles) can introduce erroneous feature points. If these are not removed, it will lead to pose estimation drift, map pollution, or even system crashes.

[0006] In recent years, researchers have proposed several improved solutions for dynamic scene problems, such as DS-SLAM, DynaSLAM, and YOLO_ORB_SLAM3. These methods typically run semantic segmentation networks (such as MaskR-CNN) or object detection models (such as YOLO) in parallel at the SLAM front end to identify dynamic categories such as "people" and "vehicles," and remove feature points in the corresponding regions before feature matching. These methods perform remarkably well in well-lit daytime scenes. However, their performance is highly dependent on the input quality of the integrated neural networks—and these networks themselves also require clear, well-lit images. Under low-light conditions, semantic segmentation accuracy drops significantly, leading to inaccurate identification of dynamic regions. This may result in the accidental deletion of static features or the retention of dynamic features, further degrading system performance.

[0007] In addition, traditional ORB feature extraction has the following inherent drawbacks:

[0008] 1) FAST corner detection uses a fixed threshold, which cannot adapt to drastic changes in lighting conditions and is prone to oversampling in bright areas and undersampling in dark areas;

[0009] 2) Binary descriptors such as BRIEF are sensitive to changes in viewpoint, lighting, and motion blur, which leads to a decrease in feature matching recall and accuracy, and seriously affects the success rate of loop closure detection and relocation. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the prior art and provide a low-light dynamic visual SLAM method and system that integrates robust features.

[0011] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0012] A low-light dynamic visual SLAM method incorporating robust features includes:

[0013] The input low-light original RGB image is subjected to non-local mean denoising to obtain the denoised image;

[0014] Determine whether the average brightness of the denoised image meets the preset conditions. If it does not meet the conditions, the denoised image is subjected to Zero-DCE brightness enhancement and CLAHE detail enhancement processing in sequence to obtain the enhanced image. If it meets the conditions, the denoised image is directly used as the enhanced image.

[0015] The enhanced image is dynamically segmented using the YOLOv8-Seg model to generate a dynamic object pixel mask.

[0016] Calculate the FAST adaptive threshold based on the average gray value of the enhanced image, and extract feature points based on the FAST adaptive threshold.

[0017] Based on the dynamic object pixel mask, feature points located within the dynamic object region are removed, while static feature points are retained;

[0018] The ORB descriptor of the static feature points is enhanced by the FeatureBooster deep neural network to generate an ORB-Booster robust descriptor.

[0019] The SLAM process is executed based on the ORB-Booster robust descriptor, including front-end tracking, local mapping, loop closure detection, and global optimization.

[0020] Furthermore, non-local mean denoising is performed on the input low-light original RGB image, including:

[0021] Convert the original RGB image to a color space with separate luminance and chrominance;

[0022] Perform the following operations on both the luminance and chroma channels of the color space:

[0023] Calculation in pixels and Euclidean distance between the central image patches , and In pixels and pixels The image patch centered on;

[0024] According to the Euclidean distance Calculate the weight function :

[0025] ;

[0026] In the luminance channel, for , represents the luminance filtering smoothing parameter; in the chroma channel, for , represents the chroma filtering smoothing parameter; It is the sum of all weights;

[0027] Perform a weighted average:

[0028] ;

[0029] In the luminance channel, for ,express The brightness value of the point; in the chroma channel, for ,express The chromaticity value of a point; In pixels The set of positions of all pixels within a finite search window centered on the target;

[0030] The denoised image is obtained.

[0031] Furthermore, Zero-DCE brightness enhancement is performed on the denoised image, including:

[0032] High-order tone curve parameters are estimated using the lightweight deep neural network DCE-Net.

[0033] ;

[0034] in, These are the original pixel values. It is a coordinate vector.

[0035] Furthermore, CLAHE detail enhancement is performed on the denoised image, including:

[0036] 1) Histogram statistics and grayscale mapping based on local regions:

[0037] The denoised image is divided into several local regions of equal size;

[0038] Perform the following operations on each local region:

[0039] The probability distribution of gray levels is obtained by statistically analyzing the frequency of pixels at each gray level within a local area. :

[0040] ;

[0041] in, grayscale The number of pixels appearing in this local area. This represents the total number of pixels in the region.

[0042] Calculate the cumulative distribution function :

[0043]

[0044] Based on the cumulative distribution function, the grayscale mapping function is used to convert the original pixel grayscale values. Mapped to enhanced grayscale values :

[0045]

[0046] in, The grayscale level of the denoised image;

[0047] 2) Histogram clipping and redistribution:

[0048] Find the clipping threshold :

[0049]

[0050] in, For gray levels, The shearing factor;

[0051] Traverse each block of the histogram Each A portion will be cut off, the sum of all cut-off pixels is calculated, and then the result is evenly added back to each block. ;

[0052] Bilinear interpolation is used between adjacent local regions to achieve a smooth transition:

[0053] Set pixel points Located in four local areas At the center, calculate the pixels respectively. The grayscale transformation values ​​that should be obtained using the cumulative distribution function (CDF) of these four blocks: ;

[0054] The bilinear interpolation formula is as follows:

[0055]

[0056]

[0057]

[0058] in, For pixels The intermediate result is obtained by horizontal interpolation of the mapping results of two adjacent local regions in the upper left and upper right directions. For pixels The intermediate result is obtained by horizontal interpolation of the mapping results of two adjacent local regions in the lower left and lower right directions. For pixels at position The final enhanced grayscale value at the location; is the horizontal distance, where The width of a local area; Vertical distance The height of a local area; The coordinates are the center points of the top-left adjacent blocks in the region to which the pixel belongs.

[0059] Furthermore, FAST adaptive threshold The calculation formula is as follows:

[0060]

[0061] in, For control coefficients, This represents the average grayscale value of the enhanced image.

[0062] Furthermore, the FeatureBooster deep neural network incorporates self-attention and cross-attention mechanisms. Self-attention captures the internal structural context of the image, while cross-attention searches for correspondences across image feature points, generating a more discriminative ORB-Booster robust descriptor.

[0063] Furthermore, to achieve the above objectives, the present invention also provides a low-light dynamic visual SLAM system with fused robust features, used to implement the above-mentioned low-light dynamic visual SLAM method with fused robust features, comprising:

[0064] The image enhancement module is used to sequentially perform noise reduction, brightness enhancement, and detail enhancement on the input low-light raw RGB image;

[0065] The dynamic object segmentation module uses the YOLOv8-Seg model to perform dynamic object detection and pixel-level segmentation on the enhanced image and generate dynamic object masks.

[0066] The adaptive feature extraction module calculates the FAST adaptive threshold based on the average gray value of the image, and extracts feature points according to the threshold.

[0067] The dynamic feature removal module is used to remove feature points located within the dynamic object region based on the dynamic object mask.

[0068] The robust feature description module uses the FeatureBooster deep neural network to enhance the ORB descriptor of static feature points, generating an ORB-Booster robust descriptor.

[0069] The SLAM core module performs front-end tracking, local mapping, loop closure detection, and global optimization based on the ORB-Booster robust descriptor.

[0070] Furthermore, the image enhancement module includes:

[0071] The nonlocal mean denoising unit is used for bilateral filtering denoising in a luminance-chrominance separated color space.

[0072] The Zero-DCE brightness enhancement unit estimates pixel-level tone curve parameters through a deep neural network to achieve non-linear dynamic range adjustment;

[0073] The CLAHE detail enhancement unit is used to improve local contrast and detail clarity in images.

[0074] Compared with existing technologies, the principles and advantages of this technical solution are as follows:

[0075] 1. Strong robustness: Through a three-level image enhancement strategy of "denoising + Zero-DCE + CLAHE", the quality of low-light images is significantly improved, providing reliable input for subsequent feature extraction.

[0076] 2. Adaptive Feature Extraction: The FAST adaptive threshold mechanism is introduced to solve the problem of unbalanced feature distribution in scenes with uneven lighting, where traditional fixed thresholds are used.

[0077] 3. Precise dynamic removal: Combining the high-precision semantic segmentation capabilities of YOLOv8-Seg, dynamic regions are accurately identified in the enhanced image, effectively preventing dynamic features from polluting the map.

[0078] 4. High matching reliability: The descriptors generated by FeatureBooster are highly invariant to illumination, viewpoint, and blur, which greatly improves the accuracy of feature matching and enhances tracking stability and loop closure detection success rate.

[0079] 5. Lightweight and practical: Zero-DCE is a lightweight network with controllable overall system computational overhead, making it suitable for embedded platforms and real-time applications. Attached Figure Description

[0080] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0081] Figure 1 This is a flowchart illustrating the principle of a low-light dynamic visual SLAM method that incorporates robust features, according to an embodiment of the present invention.

[0082] Figure 2 This is a connection block diagram of a low-light dynamic visual SLAM system that incorporates robust features, according to an embodiment of the present invention. Detailed Implementation

[0083] The present invention will be further described below with reference to specific embodiments:

[0084] like Figure 1 As shown in this embodiment, a low-light dynamic visual SLAM method incorporating robust features includes the following steps:

[0085] S1. Perform non-local mean denoising on the input low-light original RGB image to obtain the denoised image;

[0086] The specific process for this step is as follows:

[0087] Convert the original RGB image to a color space with separate luminance and chrominance;

[0088] Perform the following operations on both the luminance and chroma channels of the color space:

[0089] Calculation in pixels and Euclidean distance between the central image patches , and In pixels and pixels The image patch centered on;

[0090] According to the Euclidean distance Calculate the weight function :

[0091] ;

[0092] In the luminance channel, for , represents the luminance filtering smoothing parameter; in the chroma channel, for , represents the chroma filtering smoothing parameter; It is the sum of all weights;

[0093] Perform a weighted average:

[0094] ;

[0095] In the luminance channel, for ,express The brightness value of the point; in the chroma channel, for ,express The chromaticity value of a point; In pixels The set of positions of all pixels within a finite search window centered on the target;

[0096] The denoised image is obtained.

[0097] This step effectively suppresses color noise while maintaining color consistency and avoiding color cast.

[0098] S2. Determine whether the average brightness of the denoised image meets the preset conditions. If not, the denoised image is subjected to Zero-DCE brightness enhancement and CLAHE detail enhancement processing in sequence to obtain the enhanced image. If it meets the conditions, the denoised image is directly used as the enhanced image.

[0099] In this step, Zero-DCE brightness enhancement is performed on the denoised image, including:

[0100] High-order tone curve parameters are estimated using the lightweight deep neural network DCE-Net. Zero-DCE redefines image enhancement as an image-specific curve estimation problem using a lightweight deep neural network, DCE-Net. DCE-Net learns and estimates a set of "high-order tone curves" for an incoming low-light image. These learned curves are applied to each pixel of the original image to adjust the dynamic range of the image in a non-linear manner, thereby achieving brightness enhancement and contrast enhancement.

[0101] The image pixels are adjusted non-linearly using the following formula:

[0102] ;

[0103] in, These are the original pixel values. It is a coordinate vector. This is the core of the formula; for the brightest and darkest areas, that is... and Its adjustment amount is zero or close to zero, which can preserve the original details of these areas. For the intermediate grayscale areas, i.e. Apply the maximum enhancement effect.

[0104] In this step, CLAHE detail enhancement is performed on the denoised image, including:

[0105] 1) Histogram statistics and grayscale mapping based on local regions:

[0106] The denoised image is divided into several local regions of equal size;

[0107] Perform the following operations on each local region:

[0108] The probability distribution of gray levels is obtained by statistically analyzing the frequency of pixels at each gray level within a local area. :

[0109] ;

[0110] in, grayscale The number of pixels appearing in this local area. This represents the total number of pixels in the region.

[0111] Calculate the cumulative distribution function :

[0112]

[0113] Based on the cumulative distribution function, the grayscale mapping function is used to convert the original pixel grayscale values. Mapped to enhanced grayscale values :

[0114]

[0115] in, The grayscale level of the denoised image;

[0116] 2) Histogram clipping and redistribution:

[0117] Find the clipping threshold :

[0118]

[0119] in, For gray levels, The shearing factor;

[0120] Traverse each block of the histogram Each A portion will be cut off, the sum of all cut-off pixels is calculated, and then the result is evenly added back to each block. ;

[0121] 3) CLAHE is not based directly on global statistics of the entire image, but rather divides the image into several local regions and processes them separately. Applying it only to local regions would lead to a severe "blocking" effect. To address this, CLAHE uses bilinear interpolation of the transformation functions of its four nearest "blocks" for the final value of each pixel. The final value of each pixel is a "mixture" of the transformation rules of its four surrounding local regions; the closer a pixel is to a local region, the greater the weight of that region's transformation rule, thus achieving a smooth transition.

[0122] The process of performing bilinear interpolation between adjacent local regions to achieve a smooth transition includes:

[0123] Set pixel points Located in four local areas At the center, calculate the pixels respectively. The grayscale transformation values ​​that should be obtained using the cumulative distribution function (CDF) of these four blocks: ;

[0124] The bilinear interpolation formula is as follows:

[0125]

[0126]

[0127]

[0128] in, For pixels The intermediate result is obtained by horizontal interpolation of the mapping results of two adjacent local regions in the upper left and upper right directions. For pixels The intermediate result is obtained by horizontal interpolation of the mapping results of two adjacent local regions in the lower left and lower right directions. For pixels at position The final enhanced grayscale value at the location; is the horizontal distance, where The width of a local area; Vertical distance The height of a local area; The coordinates are the center points of the top-left adjacent blocks in the region to which the pixel belongs.

[0129] S3. Perform dynamic object segmentation on the enhanced image using the YOLOv8-Seg model to generate a dynamic object pixel mask.

[0130] S4. Calculate the FAST adaptive threshold based on the average gray value of the enhanced image, and extract feature points according to the FAST adaptive threshold;

[0131] FAST adaptive threshold The calculation formula is as follows:

[0132]

[0133] in, For control coefficients, This represents the average grayscale value of the enhanced image.

[0134] S5. Eliminate feature points located within the dynamic object region according to the dynamic object pixel mask, and retain static feature points;

[0135] S6. Enhance the ORB descriptor of the static feature points using the FeatureBooster deep neural network to generate an ORB-Booster robust descriptor.

[0136] Specifically, in this step, the FeatureBooster deep neural network includes a self-attention mechanism and a cross-attention mechanism. Self-attention captures the internal structural context of the image, while cross-attention finds the correspondence between feature points across the image, generating a more discriminative ORB-Booster robust descriptor.

[0137] Suppose we have feature points in two images, image A and image B. In self-attention, each feature point in image A "observes" all other feature points in image A, capturing the internal structure and context of the image. In cross-attention, each feature point in image A "observes" all feature points in image B, looking for potential correspondences and similarities. In this way, the network can identify feature points that "look reasonable" in terms of global geometry and strengthen their descriptors to make them more discriminative, while suppressing feature points that may be noise or blurry regions.

[0138] S7. Execute the SLAM process based on the ORB-Booster robust descriptor, including front-end tracking, local mapping, loop closure detection, and global optimization.

[0139] This invention integrates denoising and CLAHE detail enhancement with Zero-DCE, and adds Fast adaptive thresholding before feature extraction and FeatureBooster descriptor after feature point extraction to ensure that feature points contained in dynamic targets can be removed in low-light scenes with dynamic features.

[0140] In addition, this embodiment also includes, as follows Figure 2 The illustrated low-light dynamic visual SLAM system, which integrates robust features, is used to implement the aforementioned low-light dynamic visual SLAM method integrating robust features. It includes:

[0141] The image enhancement module is used to sequentially perform noise reduction, brightness enhancement, and detail enhancement on the input low-light raw RGB image;

[0142] The dynamic object segmentation module uses the YOLOv8-Seg model to perform dynamic object detection and pixel-level segmentation on the enhanced image and generate dynamic object masks.

[0143] The adaptive feature extraction module calculates the FAST adaptive threshold based on the average gray value of the image, and extracts feature points according to the threshold.

[0144] The dynamic feature removal module is used to remove feature points located within the dynamic object region based on the dynamic object mask.

[0145] The robust feature description module uses the FeatureBooster deep neural network to enhance the ORB descriptor of static feature points, generating an ORB-Booster robust descriptor.

[0146] The SLAM core module performs front-end tracking, local mapping, loop closure detection, and global optimization based on the ORB-Booster robust descriptor.

[0147] Specifically, the image enhancement module includes:

[0148] The nonlocal mean denoising unit is used for bilateral filtering denoising in a luminance-chrominance separated color space.

[0149] The Zero-DCE brightness enhancement unit estimates pixel-level tone curve parameters through a deep neural network to achieve non-linear dynamic range adjustment;

[0150] The CLAHE detail enhancement unit is used to improve local contrast and detail clarity in images.

[0151] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A low-light dynamic visual SLAM method incorporating robust features, characterized in that, include: The input low-light original RGB image is subjected to non-local mean denoising to obtain the denoised image; Determine whether the average brightness of the denoised image meets the preset conditions. If not, the denoised image is subjected to Zero-DCE brightness enhancement and CLAHE detail enhancement processing in sequence to obtain the enhanced image. If the conditions are met, the denoised image is directly used as the enhanced image; The enhanced image is dynamically segmented using the YOLOv8-Seg model to generate a dynamic object pixel mask. Calculate the FAST adaptive threshold based on the average gray value of the enhanced image, and extract feature points based on the FAST adaptive threshold. Based on the dynamic object pixel mask, feature points located within the dynamic object region are removed, while static feature points are retained; The ORB descriptor of the static feature points is enhanced by the FeatureBooster deep neural network to generate an ORB-Booster robust descriptor. The SLAM process is executed based on the ORB-Booster robust descriptor, including front-end tracking, local mapping, loop closure detection, and global optimization.

2. The low-light dynamic visual SLAM method with robust feature fusion according to claim 1, characterized in that, Non-local mean denoising is performed on the input low-light raw RGB image, including: Convert the original RGB image to a color space with luminance and chrominance separated; Perform the following operations on both the luminance and chrominance channels of the color space: Calculation in pixels and Euclidean distance between the central image patches , and In pixels and pixels The image patch centered on; According to the Euclidean distance Calculate the weight function : ； In the luminance channel, for , represents the luminance filtering smoothing parameter; in the chroma channel, for , represents the chroma filtering smoothing parameter; It is the sum of all weights; Perform a weighted average: ； In the luminance channel, for ,express The brightness value of the point; in the chroma channel, for ,express The chromaticity value of a point; In pixels The set of positions of all pixels within a finite search window centered on the target; The denoised image is obtained.

3. The low-light dynamic visual SLAM method with robust feature fusion according to claim 1, characterized in that, Zero-DCE brightness enhancement is performed on the denoised image, including: High-order tone curve parameters are estimated using the lightweight deep neural network DCE-Net. The image pixels are adjusted non-linearly using the following formula: ； in, These are the original pixel values. It is a coordinate vector.

4. The low-light dynamic visual SLAM method with robust feature fusion according to claim 1, characterized in that, CLAHE detail enhancement is performed on the denoised image, including: 1) Histogram statistics and grayscale mapping based on local regions: The denoised image is divided into several local regions of equal size; Perform the following operations on each local region: The probability distribution of gray levels is obtained by statistically analyzing the frequency of pixels at each gray level within a local area. : ； in, grayscale The number of pixels appearing in this local area, where N is the total number of pixels in the area; Calculate the cumulative distribution function : ； Based on the cumulative distribution function, the grayscale mapping function is used to convert the original pixel grayscale values. Mapped to enhanced grayscale values : ； in, The grayscale level of the denoised image; 2) Histogram clipping and redistribution: Find the clipping threshold : ； in, For gray levels, The shearing factor; Traverse each block of the histogram Each A portion will be cut off, the sum of all cut-off pixels is calculated, and then the result is evenly added back to each block. ; 3) Perform bilinear interpolation between adjacent local regions to achieve a smooth transition: Set pixel points Located in four local areas At the center, calculate the pixels respectively. The grayscale transformation values ​​that should be obtained using the cumulative distribution function (CDF) of these four blocks: ; The bilinear interpolation formula is as follows: ； ； ； in, For pixels The intermediate result is obtained by horizontal interpolation of the mapping results of two adjacent local regions in the upper left and upper right directions. For pixels The intermediate result is obtained by horizontal interpolation of the mapping results of two adjacent local regions in the lower left and lower right directions. For pixels at position The final enhanced grayscale value at the location; is the horizontal distance, where The width of a local area; Vertical distance The height of a local area; The coordinates are the center points of the top-left adjacent blocks in the region to which the pixel belongs.

5. The low-light dynamic visual SLAM method with robust feature fusion according to claim 1, characterized in that, FAST adaptive threshold The calculation formula is as follows: ； in, For control coefficients, This represents the average grayscale value of the enhanced image.

6. The low-light dynamic visual SLAM method with robust feature fusion according to claim 1, characterized in that, The FeatureBooster deep neural network incorporates self-attention and cross-attention mechanisms. Self-attention captures the internal structural context of an image, while cross-attention searches for correspondences across image feature points, generating a more discriminative ORB-Booster robust descriptor.

7. A low-light dynamic visual SLAM system incorporating robust features, used to implement the low-light dynamic visual SLAM method incorporating robust features as described in any one of claims 1-6, characterized in that, include: The image enhancement module is used to sequentially perform noise reduction, brightness enhancement, and detail enhancement on the input low-light raw RGB image; The dynamic object segmentation module uses the YOLOv8-Seg model to perform dynamic object detection and pixel-level segmentation on the enhanced image and generate dynamic object masks. The adaptive feature extraction module calculates the FAST adaptive threshold based on the average gray value of the image, and extracts feature points according to the threshold. The dynamic feature removal module is used to remove feature points located within the dynamic object region based on the dynamic object mask. The robust feature description module uses the FeatureBooster deep neural network to enhance the ORB descriptor of static feature points, generating an ORB-Booster robust descriptor. The SLAM core module performs front-end tracking, local mapping, loop closure detection, and global optimization based on the ORB-Booster robust descriptor.

8. A low-light dynamic visual SLAM system with robust features as described in claim 7, characterized in that, The image enhancement module includes: The nonlocal mean denoising unit is used for bilateral filtering denoising in a luminance-chrominance separated color space. The Zero-DCE brightness enhancement unit estimates pixel-level tone curve parameters through a deep neural network to achieve non-linear dynamic range adjustment; The CLAHE detail enhancement unit is used to improve local contrast and detail clarity in images.

Images