A Low-Light Image Enhancement Method and System for 5G Remote Control Systems Based on Deep Retinex Decomposition

Abstract

This invention relates to the field of image enhancement technology and discloses a low-light image enhancement method and system for 5G remote control systems based on deep Retinex decomposition. The method acquires and preprocesses the original low-light image from the front-end device of the 5G remote control system, inputs it into a lightweight multi-scale illumination sensing network to separate the initial illumination and reflection components, corrects the initial illumination component using a dynamic dual-branch exposure fusion module, processes the initial reflection component using a residual noise-detail co-optimization module, and then reconstructs and enhances the image through a halo and color joint correction mechanism. The image is then encoded, compressed, and transmitted to the remote control center via a 5G private network. This invention solves the core problems of underexposure and overexposure, noise interference, detail loss, and wasted bandwidth in low-light images in 5G remote control scenarios, achieving lightweight real-time enhancement at the edge, ensuring image quality and transmission efficiency, and providing reliable support for remote control decisions.

Description

Technical Field

[0001] This invention relates to the field of image enhancement technology, and in particular to a low-light image enhancement method and system for a 5G remote control system based on depth Retinex decomposition. Background Technology

[0002] 5G remote control systems provide strong support for remote real-time control due to their high bandwidth and low latency. However, the images acquired at the front end are often affected by extreme lighting conditions, including insufficient light, local strong light, and sudden changes in light. Combined with factors such as channel compression, these images exhibit complex degradation problems such as low brightness, high noise, color distortion, and loss of detail. This type of degradation severely restricts the performance of backend intelligent algorithms and the reliability of the system. Traditional Retinex-based image enhancement methods, such as MSRCR, rely on manually designed priors and exhibit poor generalization ability in complex scenes. While recent deep learning-based Retinex methods, such as RetinexNet, have achieved better decomposition results through end-to-end learning, they still suffer from significant drawbacks: First, their modules are functionally limited, lacking explicit physical constraints on illumination component estimation, leading to halo artifacts. Furthermore, their handling of reflection components fails to effectively coordinate denoising and detail enhancement, often resulting in noise amplification or detail blurring. Second, they lack adaptability; fixed network parameters cannot dynamically adjust based on image exposure levels, noise levels, scene structure, etc., resulting in insufficient robustness to the complex and variable inputs of 5G remote control systems. Finally, they are not lightweight; the original model has a large number of parameters, making direct deployment on edge devices with limited computing power difficult, failing to meet the real-time requirements of 5G remote control. Therefore, a novel network architecture that inherits the powerful fitting capabilities of deep learning while embedding physical priors and adaptive mechanisms is urgently needed to address these pain points. Summary of the Invention

[0003] To achieve the above objectives, the present invention provides the following technical solution: A low-light image enhancement method for a 5G remote control system based on deep Retinex decomposition includes: Step S1: Collect the raw low-light image data of the 5G remote control system front-end device, perform pixel value normalization processing and brightness statistical feature extraction to obtain a normalized low-light image; Step S2: Input the normalized low-light image into a lightweight multi-scale illumination sensing network for encoding and decoding processing and feature fusion to obtain the initial illumination component and the initial reflection component. Step S3: Dynamic thresholding and dual-branch differential exposure processing are performed on the initial illumination component, and the corrected illumination component is obtained through region weighted fusion. Step S4: The initial reflection component is evaluated for noise level and divided into regions. Differentiated optimization processing is performed for different noise regions to obtain the optimized reflection component. Step S5: Input the corrected illumination component and the optimized reflection component into the halo and color joint correction mechanism, and obtain the final illumination component and the final reflection component through gradient constraints and scene adaptive color correction. Step S6: Perform pixel-level multiplication on the final illumination component and the final reflection component to obtain enhanced image data, which is then encoded, compressed, and pushed to the remote control center via a 5G private network.

[0004] Furthermore, the specific process of acquiring and processing the raw low-light image data of the 5G remote control system front-end device in step S1 includes: By continuously acquiring raw low-light image data through image acquisition terminals deployed at the front-end work surface, raw low-light image data with dynamic resolution is obtained. The raw low-light image data is input into the image preprocessing unit of the edge computing device, and pixel value normalization processing is performed to obtain a normalized low-light image. The edge device memory monitoring unit detects the current available memory capacity in real time, initiates an adaptive block partitioning strategy based on memory occupancy, determines the block size, and transmits it to the network processing module.

[0005] Furthermore, the execution steps of the adaptive block partitioning strategy include: When the memory usage rate exceeds the first threshold, the normalized low-light image is segmented into image blocks of 128×128 pixels. When the memory usage rate is detected to be below the first threshold, the normalized low-light image is segmented into image blocks of 256×256 pixels.

[0006] Furthermore, the construction and processing flow of the lightweight multi-scale illumination sensing network in step S2 includes: The encoder part of the lightweight multi-scale illumination sensing network is constructed, which consists of three depthwise separable convolutional blocks stacked sequentially. The decoder part of a lightweight multi-scale illumination sensing network is constructed, which consists of three transposed convolutional blocks stacked sequentially. Perform feature fusion operation by concatenating the feature maps of the corresponding levels of the encoder and decoder along the channel dimension through a skip connection mechanism; Based on the feature map output by the decoder, the initial illumination component and the initial reflection component are generated by processing them separately through a branch structure.

[0007] Furthermore, the steps for constructing the encoder portion include: The first depthwise separable convolutional block takes a normalized low-light image as input and outputs a 16-channel downsampled feature map; The second depthwise separable convolutional block takes the feature map output from the previous block as input, uses an asymmetric convolutional kernel, and outputs a 32-channel downsampled feature map. The third depthwise separable convolutional block takes the feature map output from the previous block as input, uses an asymmetric convolutional kernel, and outputs a 64-channel downsampled feature map. Each depthwise separable convolutional block is followed by a batch normalization layer and a linear rectified activation function.

[0008] Furthermore, the dynamic threshold division and dual-branch differential exposure processing flow in step S3 includes: Calculate the global average brightness of the initial illumination component; Dynamic thresholding is performed based on the global brightness average to generate underexposure and overexposure thresholds; Create three binary mask matrices based on the generated dual thresholds; Perform dual-branch differential exposure processing, and apply corresponding processing methods to underexposed areas, overexposed areas, and normally exposed areas respectively; Perform region-weighted fusion to obtain the corrected illumination component.

[0009] Furthermore, the noise level assessment and differential optimization process in step S4 includes: Perform a noise level assessment and generate a noise intensity distribution map of the same size as the initial reflection component; Based on the noise intensity distribution map, thresholds for strong noise areas and weak noise areas are set to divide different noise regions and generate corresponding mask matrices. A nonlocal means denoising algorithm is applied to areas with strong noise. A guided filtering algorithm is applied to areas with weak noise. Apply the Laplacian sharpening algorithm to noise-free regions; The optimization results of the three regions are multiplied at the pixel level with the corresponding mask matrix, and then added and fused element by element to obtain the optimized reflection component.

[0010] Furthermore, the scene adaptive color correction step in step S5 includes: Calculate and optimize the mean vector of each color channel of the reflection component; Load the standard color channel mean vector corresponding to the current application scenario from the scene configuration library, calculate the color correction coefficient vector, and perform preliminary color balancing on the optimized reflection component; Based on the gray-scale world assumption, the overall mean of the image after preliminary correction is calculated, the fine-tuning gain coefficient is determined, and the reflection component is subjected to secondary color balance to obtain the final reflection component.

[0011] Furthermore, the pixel-level multiplication operation and encoding compression transmission process in step S6 includes: Perform pixel-level multiplication operations to construct a floating-point representation of the enhanced image data; The floating-point pixel values ​​of the enhanced image data are cropped and quantized to generate an 8-bit standard image format. The enhanced image data of consecutive frames is compressed and encoded using an efficient video coding standard; The encoded video stream is pushed to the 5G core network through the standalone networking mode of the 5G communication module, and then transmitted to the remote control center via the 5G private network. After receiving the video stream, the remote cabin control equipment decodes it, recovers the enhanced image sequence, and presents it in real time, while also outputting it to the relevant algorithm module.

[0012] A low-light image enhancement system for a 5G remote control system based on deep Retinex decomposition, used to implement the aforementioned low-light image enhancement method for a 5G remote control system based on deep Retinex decomposition, the system comprising: Image acquisition and preprocessing module: used to acquire raw low-light image data from the front-end equipment of the 5G remote control system, perform pixel value normalization processing, brightness statistical feature extraction and adaptive block division to obtain normalized low-light images; Component separation module: used to input normalized low-light images into a lightweight multi-scale illumination sensing network for encoding and decoding processing and feature fusion to obtain initial illumination components and initial reflection components; The illumination correction module is used to perform dynamic thresholding, dual-branch differential exposure processing, and region weighted fusion on the initial illumination component to obtain the corrected illumination component. The reflection optimization module is used to assess the noise level, divide the region, and perform differential optimization on the initial reflection components to obtain optimized reflection components. Joint correction module: Used to input the corrected illumination component and the optimized reflection component into the halo and color joint correction mechanism, and obtain the final illumination component and the final reflection component through gradient constraints and scene adaptive color correction; Image reconstruction and transmission module: used to perform pixel-level multiplication operations on the final illumination component and the final reflection component to obtain enhanced image data, which is then encoded, compressed, and pushed to the remote control center via a 5G private network.

[0013] Compared to existing technologies, the advantages of this invention are as follows: This invention addresses the core problems of large illumination fluctuations, uneven noise distribution, coexistence of underexposure and overexposure, and limited computing power of edge devices in low-light images in 5G remote control scenarios. It achieves precise enhancement and efficient deployment through a layered optimization technical solution. The lightweight multi-scale illumination sensing network adopts a depthwise separable convolution and asymmetric convolution kernel design, improving the inference frame rate on edge processors and completely solving the computing power bottleneck of traditional deep learning models for edge deployment, avoiding bandwidth waste and latency issues caused by the transmission of raw low-light images. The dynamic dual-branch exposure fusion module, through adaptive threshold division and differentiated processing, accurately adapts to extreme contrast scenarios, fully restoring key details such as dark safety markings and device indicator lights while effectively suppressing highlight overflow, achieving a balance between visual comfort and information integrity. The residual noise-detail co-optimization module achieves partitioned differentiated processing based on pixel-level noise evaluation, thoroughly denoising strong noise areas, smoothing edges in weak noise areas, and sharpening details in noise-free areas, thereby improving the signal-to-noise ratio and resolving the contradiction between noise suppression and detail enhancement in traditional methods. The halo and color joint correction mechanism eliminates halo artifacts at the edges of strong light through gradient constraints, while scene-adaptive color correction combined with grayscale fine-tuning ensures color fidelity across scenes, providing accurate color basis for device status judgment. Edge-side enhancement, H.265 encoding, and 5G private network transmission work together to control end-to-end latency to the millisecond level, while improving the accuracy of backend target detection and semantic segmentation, providing highly reliable and low-latency visual support for remote control decisions. Attached Figure Description

[0014] 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.

[0015] Figure 1 This is a flowchart of a low-light image enhancement method for a 5G remote control system based on deep Retinex decomposition, as described in this invention. Figure 2 This is a schematic diagram of the lightweight multi-scale illumination sensing network in an embodiment of the present invention; Figure 3 This is a schematic diagram of dynamic threshold division and dual-branch differential exposure processing in an embodiment of the present invention; Figure 4 This is a functional block diagram of a low-light image enhancement system for a 5G remote control system based on deep Retinex decomposition, as described in this invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1: Please see Figure 1 As shown, this embodiment provides a low-light image enhancement method for a 5G remote control system based on deep Retinex decomposition, including: Step S1: Collect raw low-light image data from the front-end device of the 5G remote control system, perform pixel value normalization processing and brightness statistical feature extraction to obtain a normalized low-light image.

[0018] The raw low-light image data refers to visible light image information collected by the front-end equipment of the 5G remote control system in an operating environment where the light intensity is 30% lower than the standard lighting conditions. Specifically, it is acquired by converting the light signal into a digital signal through the CMOS image sensor built into the explosion-proof camera in the mine or the camera of the inspection robot. Its original pixel value is a discrete integer value from 0 to 255. The pixel value normalization processing refers to linearly mapping the numerical range of the raw low-light image data to the floating-point range of 0 to 1. The numerical values ​​are standardized by dividing each pixel value by 255, eliminating the interference of differences in pixel value dimensions under different exposure parameters on the subsequent deep learning model.

[0019] Specifically, the operation process for collecting and processing raw low-light image data from the front-end equipment of the 5G remote control system includes the following sub-steps: Step S11: The image acquisition terminal deployed at the front-end work surface continuously performs raw low-light image data acquisition at a fixed frame rate of 30 frames per second. The exposure time of the CMOS sensor is automatically adjusted according to the ambient light intensity in each acquisition cycle to ensure that identifiable image content can still be captured under extremely low light conditions. The resolution of the acquired raw low-light image data is dynamically adjusted between 1280×720 and 1920×1080 according to the device configuration. The resolution parameter is determined by the system initialization based on the trade-off between network bandwidth and processing capability, resulting in raw low-light image data with dynamic resolution.

[0020] Step S12: Input the original low-light image data into the image preprocessing unit of the edge computing device, perform pixel value normalization processing, keep the spatial resolution of the image unchanged during the processing, and only transform the pixel values ​​point by point. By dividing each pixel value by 255, a linear mapping from integer to floating-point number is completed to obtain the normalized low-light image.

[0021] Step S13: The edge device memory monitoring unit detects the current available memory capacity in real time. When the memory occupancy rate exceeds the first threshold, an adaptive block partitioning strategy is activated to divide the normalized low-light image into 128×128 pixel blocks; otherwise, 256×256 pixel blocks are used. The block size decision result is transmitted to the subsequent network processing module in the form of block configuration parameters to guide the task partitioning of the parallel computing unit. The first threshold is set by: experimentally determining the critical value of memory occupancy rate based on the total memory capacity of the edge device and the memory required for processing 128×128 and 256×256 pixel blocks to adapt to the physical resource requirements of block processing.

[0022] The purpose and effect of this step is to address the imaging challenges posed by the drastic fluctuations in lighting, extremely wide dynamic range, and complex noise characteristics in the front-end operating environments of 5G remote control systems, such as mine tunnels and industrial inspection areas. Traditional preprocessing processes with fixed parameters cannot adapt to lighting changes at different work surfaces and time periods. This step normalizes pixel values ​​to unify images acquired under different lighting intensities to a standard numerical range, effectively avoiding numerical overflow or deep learning model input distribution drift caused by sudden changes in lighting, thus ensuring the stability of model inference. The adaptive block-based strategy adjusts the processing granularity in real time according to the resource status of edge devices, ensuring the continuity and real-time performance of the processing flow under resource-constrained conditions. This resource-aware preprocessing mechanism is particularly suitable for the actual working conditions of 5G remote control systems where the computing power of edge computing modules is greatly affected by temperature and power fluctuations, avoiding processing interruptions or data loss due to insufficient memory.

[0023] Step S2: Input the normalized low-light image into a lightweight multi-scale illumination sensing network for encoding and decoding processing and feature fusion to obtain the initial illumination component and the initial reflection component.

[0024] The lightweight multi-scale illumination sensing network refers to an encoder-decoder neural network architecture based on depthwise separable convolution and asymmetric convolution kernels, used to achieve lightweight decomposition of the illumination and reflection components of normalized low-light images. The encoding-decoding process refers to the process of extracting high-level semantic features by progressively compressing spatial resolution and expanding the number of feature channels through the encoder, and then progressively restoring spatial resolution and compressing the number of feature channels through the decoder to reconstruct the component images. The feature fusion process refers to concatenating feature maps from different levels in the encoder with corresponding feature maps from the decoder along the channel dimension through skip connections to achieve multi-scale information aggregation. The initial illumination component refers to a floating-point single-channel image describing the scene's illumination intensity distribution obtained by the lightweight multi-scale illumination sensing network decomposition, with its pixel values ​​constrained between 0 and 1 by an activation function. The initial reflection component refers to a floating-point three-channel image describing the inherent reflection characteristics of an object's surface obtained by the lightweight multi-scale illumination sensing network decomposition. Please refer to the schematic diagram of the lightweight multi-scale illumination sensing network structure. Figure 2 .

[0025] Specifically, the construction and processing flow of the lightweight multi-scale illumination sensing network includes the following sub-steps: Step S21: Construct the encoder part of the lightweight multi-scale illumination sensing network. The encoder consists of three depthwise separable convolutional blocks stacked sequentially. Each depthwise separable convolutional block contains two components: a depthwise convolutional sub-block and a pointwise convolutional sub-block. The depthwise convolutional sub-block performs spatial convolution operations independently on each input channel, with a kernel size of 3×3 and a stride of 2, used to extract spatial features within each channel. The pointwise convolutional sub-block uses a 1×1 convolutional kernel to perform cross-channel information fusion across all channels, with output channels of 16, 32, and 64, respectively, progressively improving feature representation capabilities. The first depthwise separable convolutional block accepts a normalized low-light image as input. The first deep separable convolutional block outputs a 16-channel downsampled feature map. The second deep separable convolutional block takes the 16-channel feature map from the previous block as input and uses a 5×7 asymmetric convolutional kernel to expand the horizontal receptive field to adapt to the horizontal smoothing characteristics of the illumination component, outputting a 32-channel downsampled feature map. The third deep separable convolutional block takes the 32-channel feature map from the previous block as input and uses a 7×9 asymmetric convolutional kernel to further expand the receptive field, outputting a 64-channel downsampled feature map. Each deep separable convolutional block is followed by a batch normalization layer to calibrate the mean and variance of the feature distribution, and a linear rectified activation function is used to introduce nonlinear transformation capability.

[0026] Step S22: Construct the decoder part of the lightweight multi-scale illumination sensing network. The decoder consists of three transposed convolutional blocks stacked sequentially. The transposed convolution operation restores the feature map size by learning the upsampling kernel. The first transposed convolutional block takes the 64-channel feature map output from the encoder as input and outputs a 32-channel upsampled feature map. The second transposed convolutional block takes the 32-channel feature map as input and outputs a 16-channel upsampled feature map. The third transposed convolutional block takes the 16-channel feature map as input and outputs a 3-channel upsampled feature map.

[0027] Step S23: Perform feature fusion operation. Using a skip connection mechanism, concatenate the 16-channel feature map output from the first depthwise separable convolutional block of the encoder with the input feature map to be upsampled from the second transposed convolutional block of the decoder in the channel dimension. When the spatial dimensions of the two are inconsistent, use bilinear interpolation to adjust the resolution of the encoder feature map to match the size of the decoder feature map. Concatenate the 32-channel feature map output from the second depthwise separable convolutional block of the encoder with the input feature map to be upsampled from the first transposed convolutional block of the decoder in the channel dimension. Similarly, use bilinear interpolation to handle the size mismatch problem. Concatenate the 64-channel feature map output from the third depthwise separable convolutional block of the encoder with the input feature map to be upsampled from the third transposed convolutional block of the decoder in the channel dimension.

[0028] Step S24: Based on the 3-channel feature map output by the third transposed convolutional block of the decoder, process them separately through a branch structure. One path generates the initial illumination component through the Sigmoid activation function, and the other path is directly output as the initial reflection component.

[0029] The purpose and effect of this step is that traditional RetinexNet, using standard convolution operations, results in millions of model parameters, leading to extremely low inference speeds on edge devices with limited computing power, failing to meet the real-time processing requirements of 5G remote control systems. This step uses depthwise separable convolution to reduce the multiplication operation to about one-ninth of that of standard convolution, while maintaining a relatively unchanged feature extraction capability. The asymmetric convolution kernel design expands the receptive field horizontally by 40% without increasing the number of parameters, making it more suitable for the physical characteristics of slowly changing horizontal illumination in 5G remote control scenarios, such as the horizontal extension of light bands in mine tunnels. The skip-connection channel splicing method, instead of element-wise addition, avoids aliasing loss of feature information during the fusion process, especially preserving the importance of low-level texture details for reflection component reconstruction, because the reflection component contains a large amount of high-frequency detail information, which can easily lead to detail submersion when added element-wise. After training, the number of parameters in the network is controlled within 0.038M, and a processing speed of 37fps can be achieved on the ARM Cortex-A72 processor. This fundamentally solves the deployment bottleneck of deep learning models on the 5G remote control edge, making it possible to run complex image enhancement algorithms in real time on the front-end edge device, and avoiding the bandwidth occupation and latency problems caused by transmitting the original low-light images to the cloud for processing.

[0030] For example, the initial illumination component and initial reflection component generated in step S2 can be directly used as input variables for steps S3 and S4, providing the basic components to be processed for dynamic threshold division and noise level assessment, respectively, ensuring the continuity of the variable chain in the parallel optimization process of illumination and reflection components.

[0031] Step S3: Perform dynamic thresholding and dual-branch differential exposure processing on the initial illumination component, and obtain the corrected illumination component through regional weighted fusion.

[0032] The dynamic threshold division refers to the adaptive generation of underexposure and overexposure thresholds based on the global brightness mean of the initial illumination component, used to divide image pixels into underexposed, normally exposed, and overexposed regions. The dual-branch differential exposure processing refers to a method of adjusting brightness in underexposed and overexposed regions using two different strategies: gamma correction and piecewise linear compression. The region-weighted fusion refers to the linear combination of the processing results of the three exposure regions according to preset weight coefficients to generate the final corrected illumination component. The corrected illumination component refers to the floating-point single-channel image after adaptive exposure correction, which has a more balanced brightness distribution, enhanced details in underexposed areas, and suppressed highlights in overexposed areas.

[0033] Specifically, the dynamic threshold division and dual-branch differentiated exposure processing flow includes the following sub-steps: Step S31: Calculate the global average brightness of the initial illumination component. This average is obtained by summing the brightness values ​​of all pixels in the initial illumination component and dividing by the total number of pixels, serving as a quantitative representation of the overall exposure level of the scene.

[0034] Step S32: Perform dynamic threshold division based on the global brightness mean. The underexposure threshold is composed of the sum of the products of the basic underexposure constant, the global brightness mean, and the underexposure adjustment coefficient. The overexposure threshold is composed of the difference between the products of the basic overexposure constant, the global brightness mean, and the underexposure adjustment coefficient. The basic underexposure constant, the basic overexposure constant, and the underexposure adjustment coefficient are pre-calibrated by statistically analyzing the illumination distribution characteristics of 100,000 low-light image samples collected by the 5G remote control system in mine and industrial inspection scenarios.

[0035] Step S33: Create three binary mask matrices based on the generated dual thresholds. In the underexposed area mask matrix, the corresponding positions where the pixel value is less than the underexposed threshold are set to 1, and the rest are set to 0. In the normal exposure area mask matrix, the corresponding positions where the pixel value is between the underexposed threshold and the overexposed threshold are set to 1, and the rest are set to 0. In the overexposed area mask matrix, the corresponding positions where the pixel value is greater than the overexposed threshold are set to 1, and the rest are set to 0.

[0036] Step S34: Perform dual-branch differential exposure processing. For underexposed areas, gamma correction is applied. The correction calculation involves exponentially calculating the brightness values ​​of pixels in the underexposed area. An exponent greater than 1 achieves non-linear stretching of pixels in dark areas. The gamma value is set based on the human eye's visual perception characteristics of dark areas. For overexposed areas, piecewise linear compression is performed. Pixels with values ​​less than or equal to the overexposed threshold retain their original values. Pixels with values ​​greater than the overexposed threshold are linearly mapped according to a compression slope less than 1 to suppress highlights. Normally exposed areas retain the original values ​​of the initial illumination components. Please refer to the diagram illustrating dynamic threshold division and dual-branch differential exposure processing. Figure 3 .

[0037] Step S35: Perform region weighted fusion. Multiply the underexposed region correction result by the underexposed region mask matrix and then multiply by the underexposed region weight coefficient. Multiply the normal exposure region result by the normal exposure region mask matrix and then multiply by the normal exposure region weight coefficient. Multiply the overexposed region correction result by the overexposed region mask matrix and then multiply by the overexposed region weight coefficient. Add the three weighted matrices element by element to obtain the corrected illumination component.

[0038] The purpose and effect of this step is to address the extreme contrast scenarios commonly encountered in the front-end operation environment of 5G remote control systems, such as the coexistence of intense direct headlights from vehicles in mine tunnels and absolute darkness deep within the tunnels. Traditional fixed threshold methods cannot simultaneously achieve both shadow detail recovery and highlight suppression, often resulting in either underexposure in shadows or overexposure in highlights. The dynamic threshold division mechanism adaptively adjusts the exposure area judgment standard according to the overall scene brightness. When the scene is generally dark, it automatically lowers the underexposure threshold, identifying more areas as underexposed and enhancing them. When there are strong light sources in the scene, it automatically raises the overexposure threshold, strictly controlling the highlight overflow range, thus solving the problem of insufficient adaptability caused by the rigid parameters of traditional methods. The dual-branch differentiated processing strategy is based on the physical imaging principle. It uses gamma correction that conforms to the characteristics of human vision for underexposed areas and piecewise linear compression that maintains monotonicity for overexposed areas, avoiding the loss of information in normally exposed areas caused by global uniform processing. The regional weighted fusion mechanism assigns weights to each region based on its importance to remote control decisions. Underexposed regions have higher weights to ensure full recovery of details in dark areas, which helps operators identify safety signs or equipment status indicators on the tunnel walls. Normally exposed regions have moderate weights to maintain visual naturalness, while overexposed regions have lower weights but not zero weights to retain necessary highlight information. The weight allocation strategy is determined through a combination of subjective visual evaluation and objective indicators such as peak signal-to-noise ratio and structural similarity, achieving a balance between visual comfort and information integrity. This significantly improves the readability and decision reliability of 5G remote control images under complex lighting conditions.

[0039] For example, the corrected illumination component generated in step S3 can be directly used as the input variable in step S5, and together with the optimized reflection component generated in step S4, it can be input into the halo and color joint correction mechanism to ensure that the result after illumination correction can be further processed to remove artifacts.

[0040] Step S4: Assess the noise level and divide the region for the initial reflection component, and perform differentiated optimization processing for different noise regions to obtain the optimized reflection component.

[0041] The noise level assessment refers to quantifying the noise intensity within the neighborhood of each pixel in the initial reflection component through local variance calculation, generating a noise intensity distribution map. The region division refers to dividing the initial reflection component into strong noise regions, weak noise regions, and noise-free regions based on the noise intensity distribution map. The differentiated optimization processing refers to applying three different algorithms—non-local mean denoising, guided filtering smoothing, and Laplacian sharpening—to the three noise regions respectively. The optimized reflection component refers to a floating-point three-channel image after combined noise suppression and detail enhancement processing, which improves the signal-to-noise ratio and preserves texture details.

[0042] Specifically, the noise level assessment and differential optimization process includes the following sub-steps: Step S41: Perform noise level assessment. Slide a 5×5 pixel sliding window across the initial reflection component pixel by pixel. Calculate the mean brightness of the set of pixels covered by each window. Then calculate the squared deviation of each pixel value in the window from the mean. Sum all squared deviations and divide by the total number of pixels in the window to obtain the local variance value at the center of the window. After traversing all pixel positions, generate a noise intensity distribution map of the same size as the initial reflection component.

[0043] Step S42: Based on the statistical characteristics of the noise intensity distribution map, set the thresholds for strong noise area and weak noise area. The threshold for strong noise area is obtained by adding 1.5 times the interquartile range to the upper quartile of the noise intensity distribution map. The threshold for weak noise area is obtained by calculating the median of the noise intensity distribution map. Pixels with local variance values ​​greater than the strong noise area threshold in the noise intensity distribution map are marked as strong noise areas, and a strong noise area mask matrix is ​​generated. Pixels with local variance values ​​between the weak noise area threshold and the strong noise area threshold are marked as weak noise areas, and a weak noise area mask matrix is ​​generated. Pixels with local variance values ​​less than the weak noise area threshold are marked as noise-free areas, and a noise-free area mask matrix is ​​generated.

[0044] Step S43: Apply a nonlocal mean denoising algorithm to the strong noise region. Specifically, take each pixel in the strong noise region as the center, traverse all possible center positions of similar blocks within a 7×7 search window, extract a 3×3 pixel block for each candidate position, calculate the similarity weight between the block and the block containing the center pixel. The similarity weight is obtained by calculating and normalizing the sum of squared brightness differences between the two blocks. Then, the pixel values ​​of all candidate positions are weighted and averaged according to the similarity weight. The weighted average result is used as the pixel value after denoising the strong noise region.

[0045] Step S44: Apply a guided filtering algorithm to the weak noise region. Specifically, use the initial reflection component itself as the guide image. For each pixel in the weak noise region, calculate the mean and variance of the guide image in its 5×5 neighborhood. Calculate the linear transformation coefficients based on the variance. Apply the transformation coefficients to the neighboring pixels of the input image to achieve edge-preserving smoothing.

[0046] Step S45: Apply the Laplacian sharpening algorithm to the noise-free region. Specifically, perform convolution operation on each pixel in the noise-free region with its eight neighboring pixels. The center weight of the convolution kernel is 5, the weight of the four-connected neighbor is -1, and the weight of the diagonal neighbor is 0 to enhance high-frequency details.

[0047] Step S46: Perform pixel-level multiplication operations on the optimization results of the three regions with the corresponding mask matrices to extract the processing results of each region. Then, add the results of the three regions element by element and fuse them to obtain the optimized reflection component.

[0048] The purpose and effect of this step is to address the significant spatial non-uniformity of noise distribution in low-light images acquired by 5G remote control systems. Dark areas exhibit strong noise due to weak signals, while bright areas have weak noise due to strong signals. Traditional methods using uniform denoising intensity lead to over-smoothing in weak noise areas, resulting in the loss of subtle textures, while incomplete denoising in strong noise areas leaves residual particle noise. This step's noise level estimator achieves pixel-level precise quantification of noise intensity, dividing the image into regions with different noise levels. This solves the problem that a single denoising parameter cannot adapt to spatially variable noise. This division method fully considers the noise non-uniformity caused by differences in reflectivity of different object materials in 5G remote control scenarios; for example, metal surfaces reflect light strongly but have weak noise, while coal walls absorb light strongly and have strong noise. The non-local mean algorithm utilizes image self-similarity in strong noise areas, effectively distinguishing random noise from recurring real textures by searching for structural blocks similar to the current block within a large search window. This avoids the detail blurring caused by traditional local averaging filtering, making it particularly suitable for denoising repetitive structures such as tunnel support components and equipment shells in 5G remote control scenarios. Guided filtering, through guided graph constraints, ensures the smoothing process follows the object's edge direction in low-noise regions. This removes minor noise while preserving the integrity of the object's contour, preventing edge blurring from affecting subsequent object detection algorithms. Laplacian sharpening specifically enhances detail contrast in noise-free regions, avoiding noise amplification caused by sharpening in noisy areas. This achieves synergistic optimization of noise suppression and detail enhancement, resolving the contradiction of traditional methods that treat these two aspects as opposing forces.

[0049] For example, the optimized reflection component generated in step S4 can be directly used as the input variable in step S5, and together with the corrected illumination component generated in step S3, it can be input into the halo and color joint correction mechanism to ensure that the result of the optimized reflection component can be further processed by color correction.

[0050] Step S5: Input the corrected illumination component and the optimized reflection component into the halo and color joint correction mechanism, and obtain the final illumination component and the final reflection component through gradient constraints and scene adaptive color correction.

[0051] The halo and color joint correction mechanism refers to a joint optimization framework that simultaneously handles halo artifacts in the illumination component and color distortion in the reflection component. The gradient constraint refers to suppressing halos by constraining the difference between the gradient distribution of the illumination component and the gradient distribution after Gaussian smoothing. The scene-adaptive color correction refers to performing color balancing on the optimized reflection component based on the preset standard color mean of the 5G remote control system application scenario. The final illumination component refers to the floating-point single-channel image after halo suppression. The final reflection component refers to the floating-point three-channel image after color correction.

[0052] Specifically, the processing flow of the halo and color joint correction mechanism includes the following sub-steps: Step S51: Perform halo suppression processing and construct a corrected illumination loss function. First, calculate the horizontal gradient map and vertical gradient map of the corrected illumination component. The horizontal gradient is calculated by the brightness difference between each pixel and the pixel to its right, and the vertical gradient is calculated by the brightness difference between each pixel and the pixel below it.

[0053] Step S52: Perform Gaussian filtering smoothing on the corrected illumination component with a Gaussian kernel standard deviation of 1.5. After smoothing, recalculate the horizontal and vertical gradient maps, calculate the absolute difference between the original gradient map and the smoothed gradient map, and use the mean of this difference as the halo suppression term.

[0054] Step S53: Combine the smoothing loss of the corrected illumination component with the halo suppression term according to the weight coefficients to form the corrected illumination optimization objective. Iteratively optimize the corrected illumination component using the gradient descent algorithm to make the gradient distribution in the illumination change region approach smoothness and eliminate local overshoot. During the iterative optimization process, the learning rate is dynamically adjusted according to the optimization stability to obtain the final illumination component.

[0055] Step S54: Perform scene adaptive color correction. First, calculate the mean vector of the optimized reflection component in the red, green and blue channels. The mean of each channel is obtained by summing all pixel values ​​of that channel and dividing by the total number of pixels.

[0056] Step S55: Load the standard color channel mean vector corresponding to the current application scenario from the scenario configuration library of the 5G remote control system. The standard color mean is obtained by collecting thousands of normal lighting reference images under the scenario and calculating the channel mean. Calculate the ratio of the standard color mean vector to the current image channel mean vector to obtain the color correction coefficient vector. Multiply all pixel values ​​of each channel of the optimized reflection component by the color correction coefficient of the corresponding channel to achieve preliminary color balance.

[0057] Step S56: Based on the gray-scale world assumption, calculate the overall mean of the image after preliminary correction. This overall mean is obtained by summing the mean values ​​of the three channels and dividing by 3. Use the overall mean as the target mean. Calculate the ratio of the mean value of each current channel to the target mean to obtain the fine-tuning gain coefficient. Multiply the pixel values ​​of each channel of the preliminary corrected reflection component by the corresponding fine-tuning gain coefficient to make the mean values ​​of the three channels approach the target mean value, complete the secondary color balance, and obtain the final reflection component.

[0058] The purpose and effect of this step is that the traditional Retinex method often produces halo artifacts at the edges of strong light after adjusting the illumination components. These artifacts appear as white halos formed by brightness overshoot, severely affecting image visual quality and target recognition accuracy. The gradient constraint mechanism innovatively unifies the modeling of illumination smoothness and halo suppression. By comparing the difference between the original gradient and the smoothed gradient, it automatically identifies and penalizes abnormal transition regions, making the illumination transition conform to the gradual illumination physical characteristics of natural objects. This fundamentally eliminates the mathematical basis for halo generation and avoids the additional computational overhead of post-processing halo removal methods. 5G remote control systems have diverse application scenarios. Mine scenes present a reddish-brown tone due to the reflective properties of coal dust and rocks, while industrial inspection scenes present a bluish-gray tone due to metal equipment and fluorescent lighting. Fixed color correction matrices cannot adapt to the color constancy requirements across different scenarios. Scene adaptive color correction incorporates prior scene knowledge into the correction process by loading a preset scene standard color mean. This solves the problem of automatic white balance failure caused by insufficient illumination, ensuring that the image colors seen by remote operators are consistent with the actual colors on site. This is beneficial for judging key information such as the color of indicator lights and the color of liquid leaks. The secondary correction mechanism of grayscale world fine-tuning further compensates for the channel imbalance that may be introduced by the initial correction, ensuring that the color fidelity of the final reflection component meets the visual accuracy requirements of remote control decision-making.

[0059] For example, the final illumination component and the final reflection component generated in step S5 can be directly used as input variables in step S6. Image reconstruction is completed through pixel-level multiplication operations to ensure that the corrected and optimized components can be finally integrated into an enhanced image.

[0060] Step S6: Perform pixel-level multiplication on the final illumination component and the final reflection component to obtain enhanced image data, which is then encoded, compressed, and pushed to the remote control center via a 5G private network.

[0061] The pixel-level multiplication operation refers to an image reconstruction operation that performs element-wise multiplication of each pixel value of the final illumination component with the corresponding pixel value of the final reflection component. The enhanced image data refers to a floating-point three-channel image generated after the pixel-level multiplication operation, whose dynamic range has been expanded and whose details and colors have been enhanced. The encoding compression refers to using video coding standards to remove redundant information and reduce the data volume of the enhanced image data. The 5G private network refers to an end-to-end network slice configured for the 5G remote control system in 5G standalone networking mode, providing guaranteed bandwidth and latency services.

[0062] Specifically, the pixel-level multiplication operation and encoding compression transmission process includes the following sub-steps: Step S61: Perform pixel-level multiplication in the image reconstruction unit of the edge device, using the final illumination component as the multiplier and the final reflection component as the multiplicand, and perform element-wise multiplication on the three color channels respectively. The multiplication result constitutes the floating-point representation of the enhanced image data. This operation realizes the Retinex reconstruction theory of illumination and reflection components.

[0063] Step S62: Crop the floating-point pixel value range of the enhanced image data to between 0 and 1, set values ​​less than 0 to 0, and values ​​greater than 1 to 1 to complete the numerical truncation process. Map the truncated floating-point image data back to the integer range of 0 to 255 through linear quantization. The mapping method is to multiply each pixel value by 255 and then round it to the nearest integer to generate an 8-bit standard image format.

[0064] Step S63: At the edge device, the H.265 video coding standard is used to compress and encode the enhanced image data of consecutive frames. The encoding process uses temporal prediction technology to predict the current frame by referring to the previous frame image, performs discrete cosine transform and quantization on the prediction residual, and then further compresses it through entropy coding to compress the data volume to less than 10% of the original size.

[0065] Step S64: The encoded video stream is pushed to the 5G core network through the standalone networking mode of the 5G communication module, and then transmitted to the remote control center through the 5G private network's quality of service guarantee mechanism. The quality of service guarantee mechanism assigns a high priority level to the video stream to ensure stable transmission even when the network is congested.

[0066] Step S65: After receiving the video stream, the remote cabin control equipment performs H.265 decoding to recover the enhanced image sequence, which is then displayed in real time on a high-definition display. The decoding process is the reverse of the encoding process. Image frames are reconstructed through entropy decoding, inverse quantization, inverse discrete cosine transform, and motion compensation. The decoded image data is synchronously output to the target detection algorithm module and the semantic segmentation algorithm module as the input source for the intelligent analysis algorithm, providing visual information support for remote control decision-making.

[0067] The purpose and effect of this step is that traditional 5G remote control systems directly transmit raw low-light images collected at the front end to the remote end. These unprocessed low-light images contain a large amount of visual information redundancy and noise interference, consuming significant 5G network bandwidth resources. Furthermore, the backend intelligent algorithm directly processes degraded images, leading to a substantial increase in false detection and false negative rates, affecting the security of remote control. This step moves image enhancement processing to the front-end edge, transmitting only the enhanced, high-quality image, significantly reducing the amount of data transmitted. This frees up 5G network bandwidth resources for transmitting critical information such as control signaling and sensor data, improving network resource utilization efficiency. The H.265 encoding at the edge end, combined with the 5G private network standalone networking mode, leverages the low latency of 5G to control end-to-end latency to the millisecond level, meeting the latency requirements of remote real-time control. The enhanced image is directly input into the backend intelligent algorithm, avoiding performance degradation caused by image degradation, improving the accuracy of target detection and semantic segmentation, and enabling remote operators to make accurate decisions based on clear visual information. This architecture design, which combines edge intelligence with 5G transmission, enables a reasonable allocation of computing tasks across the edge and cloud. The front end is responsible for real-time enhancement, while the cloud is responsible for policy optimization. This is a key technological breakthrough in the evolution of 5G remote control systems from simple data transmission to intelligent information transmission, laying the foundation for building a highly reliable, low-latency, and intelligent remote control system.

[0068] For example, after the enhanced image data generated in step S6 is transmitted to the remote control center via the 5G private network, it can be directly presented to the remote operator on a high-definition display in real time, and simultaneously input into the target detection and semantic segmentation algorithm module to complete the end-to-end process of low-light image enhancement of the entire 5G remote control system, realizing a complete technical closed loop from image acquisition, enhancement processing to remote application.

[0069] Example 2: This embodiment, based on Embodiment 1, provides a low-light image enhancement system for a 5G remote control system based on deep Retinex decomposition, such as... Figure 4 As shown, it includes: Image acquisition and preprocessing module: used to acquire raw low-light image data from the front-end equipment of the 5G remote control system, perform pixel value normalization processing, brightness statistical feature extraction and adaptive block division to obtain normalized low-light images; Component separation module: used to input normalized low-light images into a lightweight multi-scale illumination sensing network for encoding and decoding processing and feature fusion to obtain initial illumination components and initial reflection components; The illumination correction module is used to perform dynamic thresholding, dual-branch differential exposure processing, and region weighted fusion on the initial illumination component to obtain the corrected illumination component. The reflection optimization module is used to assess the noise level, divide the region, and perform differential optimization on the initial reflection components to obtain optimized reflection components. Joint correction module: Used to input the corrected illumination component and the optimized reflection component into the halo and color joint correction mechanism, and obtain the final illumination component and the final reflection component through gradient constraints and scene adaptive color correction; Image reconstruction and transmission module: used to perform pixel-level multiplication operations on the final illumination component and the final reflection component to obtain enhanced image data, which is then encoded, compressed, and pushed to the remote control center via a 5G private network.

Claims

1. A low-light image enhancement method for a 5G remote control system based on deep Retinex decomposition, characterized in that, The method includes: S1: Collect raw low-light image data from the front-end device of the 5G remote control system, perform pixel value normalization processing and brightness statistical feature extraction to obtain a normalized low-light image; S2: Input the normalized low-light image into a lightweight multi-scale illumination sensing network for encoding and decoding processing and feature fusion to obtain the initial illumination component and the initial reflection component; S3: Dynamic thresholding and dual-branch differential exposure processing are performed on the initial illumination component, and the corrected illumination component is obtained through regional weighted fusion; S4: The initial reflection component is evaluated for noise level and divided into regions. Differentiated optimization processing is performed for different noise regions to obtain the optimized reflection component. S5: Input the corrected illumination component and the optimized reflection component into the halo and color joint correction mechanism, and obtain the final illumination component and the final reflection component through gradient constraints and scene adaptive color correction; S6: Perform pixel-level multiplication on the final illumination component and the final reflection component to obtain enhanced image data, which is then encoded, compressed, and pushed to the remote control center via a 5G private network.

2. The low-light image enhancement method for a 5G remote control system based on deep Retinex decomposition according to claim 1, characterized in that, The steps for acquiring and processing raw low-light image data from the front-end equipment of the 5G remote control system include: By continuously acquiring raw low-light image data through image acquisition terminals deployed at the front-end work surface, raw low-light image data with dynamic resolution is obtained. The raw low-light image data is input into the image preprocessing unit of the edge computing device, and pixel value normalization processing is performed to obtain a normalized low-light image. The edge device memory monitoring unit detects the current available memory capacity in real time, initiates an adaptive block partitioning strategy based on memory occupancy, determines the block size, and transmits it to the network processing module.

3. The low-light image enhancement method for a 5G remote control system based on deep Retinex decomposition according to claim 2, characterized in that, The execution steps of the adaptive block partitioning strategy include: When the memory usage rate exceeds the first threshold, the normalized low-light image is segmented into image blocks of 128×128 pixels. When the memory usage rate is detected to be below the first threshold, the normalized low-light image is segmented into image blocks of 256×256 pixels.

4. The low-light image enhancement method for a 5G remote control system based on deep Retinex decomposition according to claim 1, characterized in that, The construction and processing flow of a lightweight multi-scale illumination sensing network includes: The encoder part of the lightweight multi-scale illumination sensing network is constructed, which consists of three depthwise separable convolutional blocks stacked sequentially. The decoder part of a lightweight multi-scale illumination sensing network is constructed, which consists of three transposed convolutional blocks stacked sequentially. Perform feature fusion operation by concatenating the feature maps of the corresponding levels of the encoder and decoder along the channel dimension through a skip connection mechanism; Based on the feature map output by the decoder, the initial illumination component and the initial reflection component are generated by processing them separately through a branch structure.

5. The low-light image enhancement method for a 5G remote control system based on deep Retinex decomposition according to claim 4, characterized in that, The steps involved in building the encoder section include: The first depthwise separable convolutional block takes a normalized low-light image as input and outputs a 16-channel downsampled feature map; The second depthwise separable convolutional block takes the feature map output from the previous block as input, uses an asymmetric convolutional kernel, and outputs a 32-channel downsampled feature map. The third depthwise separable convolutional block takes the feature map output from the previous block as input, uses an asymmetric convolutional kernel, and outputs a 64-channel downsampled feature map. Each depthwise separable convolutional block is followed by a batch normalization layer and a linear rectified activation function.

6. The low-light image enhancement method for a 5G remote control system based on deep Retinex decomposition according to claim 1, characterized in that, The dynamic thresholding and dual-branch differential exposure processing workflow includes: Calculate the global average brightness of the initial illumination component; Dynamic thresholding is performed based on the global brightness average to generate underexposure and overexposure thresholds; Three binary mask matrices are created based on the generated underexposure and overexposure thresholds; Perform dual-branch differential exposure processing, and apply corresponding processing methods to underexposed areas, overexposed areas, and normally exposed areas respectively; Perform region-weighted fusion to obtain the corrected illumination component.

7. The low-light image enhancement method for a 5G remote control system based on deep Retinex decomposition according to claim 1, characterized in that, The noise level assessment and differential optimization process includes: Perform a noise level assessment and generate a noise intensity distribution map of the same size as the initial reflection component; Based on the noise intensity distribution map, thresholds for strong noise areas and weak noise areas are set to divide different noise regions and generate corresponding mask matrices. A nonlocal means denoising algorithm is applied to areas with strong noise. A guided filtering algorithm is applied to areas with weak noise. Apply the Laplacian sharpening algorithm to noise-free regions; The optimization results of the three regions are multiplied at the pixel level with the corresponding mask matrix, and then added and fused element by element to obtain the optimized reflection component.

8. The low-light image enhancement method for a 5G remote control system based on deep Retinex decomposition according to claim 1, characterized in that, The steps of scene-adaptive color correction include: Calculate and optimize the mean vector of each color channel of the reflection component; Load the standard color channel mean vector corresponding to the current application scenario from the scene configuration library, calculate the color correction coefficient vector, and perform preliminary color balancing on the optimized reflection component; Based on the gray-scale world assumption, the overall mean of the image after preliminary correction is calculated, the fine-tuning gain coefficient is determined, and the reflection component is subjected to secondary color balance to obtain the final reflection component.

9. A low-light image enhancement method for a 5G remote control system based on deep Retinex decomposition according to claim 1, characterized in that, The pixel-level multiplication operation and encoding compression transmission process includes: Perform pixel-level multiplication operations to construct a floating-point representation of the enhanced image data; The floating-point pixel values ​​of the enhanced image data are cropped and quantized to generate an 8-bit standard image format. The enhanced image data of consecutive frames is compressed and encoded using an efficient video coding standard; The encoded video stream is pushed to the 5G core network through the standalone networking mode of the 5G communication module, and then transmitted to the remote control center via the 5G private network. After receiving the video stream, the remote cabin control equipment decodes it, recovers the enhanced image sequence, and presents it in real time, while also outputting it to the relevant algorithm module.

10. A low-light image enhancement system for a 5G remote control system based on deep Retinex decomposition, used to implement the method according to any one of claims 1-9, characterized in that, The system includes: Image acquisition and preprocessing module: used to acquire raw low-light image data from the front-end equipment of the 5G remote control system, perform pixel value normalization processing, brightness statistical feature extraction and adaptive block division to obtain normalized low-light images; Component separation module: used to input normalized low-light images into a lightweight multi-scale illumination sensing network for encoding and decoding processing and feature fusion to obtain initial illumination components and initial reflection components; The illumination correction module is used to perform dynamic thresholding, dual-branch differential exposure processing, and region weighted fusion on the initial illumination component to obtain the corrected illumination component. The reflection optimization module is used to assess the noise level, divide the region, and perform differential optimization on the initial reflection components to obtain optimized reflection components. Joint correction module: Used to input the corrected illumination component and the optimized reflection component into the halo and color joint correction mechanism, and obtain the final illumination component and the final reflection component through gradient constraints and scene adaptive color correction; Image reconstruction and transmission module: used to perform pixel-level multiplication operations on the final illumination component and the final reflection component to obtain enhanced image data, which is then encoded, compressed, and pushed to the remote control center via a 5G private network.

Images