A visual perception and regulation method suitable for wheel-legged robot patrol

Abstract

The application provides a visual perception and regulation method suitable for wheel-foot robot inspection, and belongs to the technical field of image processing. The original low-light image of industrial equipment is collected; the original low-light image is input into a low-light image enhancement model for image enhancement; the low-light image enhancement model processing process comprises: inputting the image into a U-net network to predict noise; denoising the original image based on the predicted noise to obtain a reconstructed image as an enhanced image; mapping the reconstructed image from a spatial domain to a frequency domain to obtain a frequency domain representation; frequency division is performed on the frequency domain representation, and a learnable weight is allocated to each frequency band, and a frequency domain constraint term is constructed based on the weighted result. In the diffusion model inversion stage, an uncertainty network is introduced, the noise branch and the residual branch weight are adaptively adjusted according to the pixel-level uncertainty and the time step, the noise in the high-noise area is strengthened, the structure in the detail area is reserved, the detail and brightness balance of the enhanced image is balanced, and the natural smoothness is output; and the dynamic balance of different regions and stages is realized.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a visual perception and control method applicable to the inspection of wheeled robots. Background Technology

[0002] Belt conveyors, as core transportation equipment, are widely used in industrial settings such as mining, metallurgy, chemical industry, and ports. Due to common environmental issues such as insufficient lighting, uneven illumination, dust dispersion, and equipment obstruction, inspection images often exhibit low brightness, poor contrast, significant noise, and loss of detail. This directly impacts the accuracy of identifying critical faults such as deviation, tearing, and spillage, as well as the stability of the system. In recent years, mobile robots, with their autonomous navigation and multi-sensor fusion capabilities, have been gradually applied to industrial inspection. However, their onboard vision systems still face image quality degradation under low light and non-uniform lighting conditions, becoming one of the main bottlenecks restricting the improvement of intelligent inspection efficiency.

[0003] In the field of low-light image enhancement, existing technologies can be mainly divided into three categories: The first category consists of traditional image processing methods based on grayscale transformation and histogram equalization. Although these methods are computationally efficient, they are prone to over-enhancement, color distortion, and noise amplification, making them unsuitable for complex industrial scenarios. The second category consists of methods based on Retinex theory, which improve image quality through illumination-reflection decomposition models. However, these methods lack robustness to non-uniform illumination, and there is a contradiction between detail restoration and noise suppression. The third category consists of deep learning methods, which achieve non-linear mapping through end-to-end training. While these methods outperform traditional methods in general scenarios, they still suffer from weak cross-domain generalization ability, poor adaptability to non-uniform illumination, and high edge computing resource consumption in specific industrial inspection scenarios. Specifically, existing technologies face the following core defects: First, the protective cover of the conveyor belt and the shadows of the equipment cause significant differences in local brightness and darkness in the image, making global enhancement prone to overexposure in bright areas and loss of detail in dark areas. Second, frequency domain noise caused by dust scattering is amplified synchronously during the enhancement process, resulting in texture artifacts. Third, the platform's computing power is limited, and existing complex networks cannot meet the real-time processing requirements.

[0004] Therefore, there is an urgent need for a solution that balances enhancement, noise suppression, and computational efficiency to support reliable visual perception in industrial inspection scenarios. Summary of the Invention

[0005] In view of this, the present invention provides a visual perception and control method applicable to the inspection of wheeled robots. In the diffusion stage, an uncertainty network is used to update the dynamic ratio of noise; in the image reconstruction stage, a lightweight U-net network is used to predict noise, thereby ensuring image quality while significantly reducing computational complexity and inference latency, which is suitable for the inspection needs of industrial equipment.

[0006] Therefore, the present invention provides the following technical solution:

[0007] A visual perception and control method applicable to inspection of wheeled robots includes:

[0008] Acquire raw, low-light images of industrial equipment;

[0009] The original low-light image is input into the low-light image enhancement model for image enhancement.

[0010] The low-light image enhancement model processing procedure includes:

[0011] The image is input into the U-net network to predict noise;

[0012] The original image is denoised based on the predicted noise to obtain a reconstructed image as the enhanced image.

[0013] The reconstructed image is mapped from the spatial domain to the frequency domain to obtain a frequency domain representation; the frequency domain representation is divided into frequencies, and learnable weights are assigned to each frequency band; frequency domain constraint terms are constructed based on the weighted results.

[0014] Furthermore, the U-net network includes:

[0015] The convolutional residual blocks in the feature extraction backbone between the encoder and decoder of the U-net network are replaced by a lightweight kernel activation function module.

[0016] The lightweight kernel activation function module includes:

[0017] Channel selection and nonlinear mapping are performed on local detail features to obtain a feature result that fuses linear and nonlinear outputs, expressed by the formula:

[0018]

[0019] in, To integrate the characteristic results of linear and nonlinear outputs; Select a mask for the channel, only for Nonlinear mapping is applied to the channels to reduce computational costs; This is an element-wise multiplication operation; It provides learnable nonlinear transformation capabilities for KAN kernel function mapping; when When, the corresponding channel undergoes a KAN nonlinear transformation; when At that time, the channel remains linear;

[0020] Wherein, the KAN kernel function is:

[0021]

[0022] in, These are the feature values ​​of the input channel; For the first Segment spline weights; The segment node position; The number of segments; Linear rectifier functions enable mappings to exhibit nonlinearity and sparsity.

[0023] Further, the step of denoising the original image based on the predicted noise to obtain the reconstructed image includes:

[0024]

[0025] in, For the first Step-by-step reconstruction of the image; The adaptive scheduling weights for residual branches; This represents the noise ratio of the current step. To predict noise; To predict residuals.

[0026] Further, the frequency domain image is divided into frequency bands and learnable weights are assigned to each band, including:

[0027] The loss weights for each frequency band are dynamically adjusted using an attention mask, as expressed by the formula:

[0028]

[0029]

[0030] in, This represents the channel average pooling characteristic. This represents the channel max pooling feature; All are learnable parameters. It is the weight matrix of the convolution. These are the corresponding bias terms; different Corresponding to different frequency bands; For mapping functions; It is a learnable spatial mask used to dynamically adjust the loss weights for each frequency band. In frequency domain constraints, It is not directly included in the loss calculation.

[0031] Furthermore, the training process of the low-light image enhancement model includes:

[0032] The process of obtaining the reconstructed image described in the first stage of training;

[0033] The loss function for the first stage is:

[0034]

[0035] in, For pixel reconstruction items, For structure preservation terms; For frequency domain constraints;

[0036] The second stage trains the overall low-light image enhancement model, and the loss function for the second stage is:

[0037]

[0038] in: Image reconstruction loss; For frequency domain constraints; Regularization is guided by uncertainty.

[0039] Furthermore, the frequency domain constraint:

[0040]

[0041] in, Low / medium / high frequency bands; For the reason Dynamically adjusted bandwidth weights; These are phase balance parameters; The amplitude spectrum contains brightness and contrast information; The phase spectrum contains structural and edge information; The diffusion model is represented in the first... -1 step generated reconstructed image; Indicates a reference to the original image; 、 : respectively the first The amplitude spectrum of an image in each frequency band reflects the brightness and contrast distribution information. 、 : respectively the first The phase spectrum of an image in each frequency band reflects structural, edge, and texture information; Spatial attention mask Through linear mapping function The generated frequency band weighting coefficients are used to dynamically adjust the contribution of each frequency band to the loss; This is the phase balance parameter, used to control the relative weights of the amplitude and phase losses. express Norm, used to measure the absolute distance of spectral differences.

[0042] Furthermore, the uncertainty guides regularization:

[0043]

[0044] Among them, if In regions where high noise levels lead to model unreliability, the regularization term has a large weight, forcing the predicted noise to match the actual noise; if... Reduce constraints in regions where the model has already stabilized to prevent overfitting.

[0045] Furthermore, the image reconstruction loss is:

[0046]

[0047] in, The reconstructed image obtained from model inversion; The original low-light input image; For the first The image after adding noise; Output for residual branches; This is a pixel-level uncertainty map for the corresponding time step; The variance of the uncertainty plot is used to measure the dispersion of prediction confidence. These are the respective balancing weights, used to coordinate pixel consistency, residual consistency, and uncertainty stability; This is a pixel consistency term to ensure that the brightness of the inverted image is consistent with that of the real image; As a residual consistency term, it constrains the model's predicted residuals to be close to the actual residuals, thereby improving the ability to reconstruct structural details. As an uncertainty stabilizing term, it suppresses the variance of the pixel-level uncertainty map and ensures the stability of the sampling process.

[0048] Furthermore, the structure-preserving item includes:

[0049]

[0050] in, This is a structure-preserving term used to maintain the consistency of the edges and structure of the original image during the enhancement process; This represents the gradient map of the reconstructed image during the diffusion inversion stage. This represents the gradient map of the input low-light image; :for Norm, used to measure gradient difference.

[0051] Advantages and positive effects of the present invention:

[0052] 1) This method introduces an uncertainty network in the diffusion model inversion stage. Based on pixel-level uncertainty and time step, it adaptively adjusts the weights of noise branches and residual branches, strengthens denoising in high-noise areas, and preserves structure in detail areas. It effectively reduces the number of diffusion steps and improves sampling stability. It enhances the balance of image details and brightness, and outputs a natural and smooth image. It achieves dynamic balance in different regions and stages.

[0053] 2) In the image reconstruction stage, this method predicts noise through the U-net network and achieves lightweight and efficient feature mapping through a lightweight kernel activation function module, which significantly reduces computational complexity and inference latency while ensuring image quality.

[0054] 3) This method performs fast Fourier transform on the reconstructed image to achieve spectral decomposition, and combines frequency band weighting and attention fusion mechanism to achieve adaptive separation of illumination and texture frequency bands, thereby enhancing edge details and suppressing noise, significantly improving brightness and color consistency, and eliminating halo artifacts. Attached Figure Description

[0055] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart of the visual perception and control method applicable to the inspection of wheeled robots in this embodiment of the invention;

[0057] Figure 2 This is a framework diagram of a low-light image enhancement model in an embodiment of the present invention;

[0058] Figure 3 This is a framework diagram of the lightweight kernel activation function module in an embodiment of the present invention;

[0059] Figure 4 This is a partial KAN framework diagram in the lightweight kernel activation function module of this invention. Detailed Implementation

[0060] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. 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 should fall within the scope of protection of the present invention.

[0061] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0062] This invention provides a visual perception and control method applicable to the inspection of wheeled robots. Combining uncertainty-driven dual-schedule sampling, adaptive frequency domain constraints, and a lightweight KAN-Lite module-based deep diffusion enhancement network, it achieves high-quality enhancement of low-light images, effectively improving the visual perception and fault identification capabilities of inspection tasks. By combining frequency domain perception and amplitude-phase consistency constraints, adaptive weighting of low / medium / high-frequency information can effectively suppress noise and artifacts, restore texture and color, and significantly improve the quality of inspection images under complex conditions such as non-uniform lighting and dust interference. This provides stable support for downstream fault detection tasks such as deviation, tearing, and material spillage.

[0063] Combination Figure 1 As shown, a visual perception and control method applicable to the inspection of wheeled robots includes:

[0064] S1, The image enhancement model receives images acquired in real time.

[0065] Low-light images were collected from multiple angles in the belt conveyor channel, idler rollers, and return belt area.

[0066] S2. Enhance the low-light image using an image enhancement model;

[0067] The image enhancement model data processing procedure includes:

[0068] 1. Perform forward diffusion on the original low-light image:

[0069]

[0070] in, Original low-light image; For the first The noisy input image of the step; The noise term follows a standard normal distribution; This is the cumulative noise attenuation coefficient. For the first The step noise ratio parameter controls the noise intensity at each step.

[0071] 2. Calculate pixel-level weights using an uncertain network. The noise is updated and allocated with a dynamic ratio to achieve adaptive regularization;

[0072] Current noisy image With time step They are fed together into a lightweight convolutional subnet, and the output is... .

[0073] Interaction function of time and uncertainty

[0074] The weights are determined by both the time step and the uncertainty:

[0075]

[0076] in, The activation function is Sigmoid, ensuring the output range is (0,1). This is an interaction function between time and uncertainty. Defined as:

[0077]

[0078] in, These are the time-step modulation coefficients; For time-aware functions. When When the value is large, it indicates prediction uncertainty and strong noise, and the model tends to update the residuals; when When the noise level is low, the model tends to suppress noise.

[0079] in, These are learnable parameters, referring to the weights and biases automatically optimized during model training using the backpropagation algorithm. They control the strength and offset of the linear mapping between the time and uncertainty interaction function. These parameters are random values ​​during model initialization and are updated based on the gradient of the loss function L after multiple training iterations to minimize training error. Therefore, Control time step versus uncertainty graph The extent of the impact Adjusting the center position of the mapping to achieve the interaction function of time and uncertainty. Its adaptive adjustment characteristics.

[0080] 3. Predicting noise terms using the U-net network With structural residuals ;

[0081] The output of the U-Net network consists of multi-scale feature maps and predicted noise terms from the enhancement stage. .

[0082] Among them, the multi-scale feature map is output through the decoding end of U-Net to provide local and global structural information, providing basic feature support for the subsequent frequency domain constraint module and uncertainty regularization module;

[0083] Predicted noise term Indicates the network in the 1st... The estimation results of noise components in the step diffusion inversion are used to guide noise removal and brightness reconstruction.

[0084] Predicting Noise Using U-Net We combine depthwise separable convolution with spline kernel functions as a lightweight kernel activation function module to achieve lightweight and efficient feature mapping.

[0085] The data processing procedure of the lightweight kernel activation function module is as follows:

[0086] Input feature map The number of channels is Space dimensions are .

[0087] 1) Perform feature dimensionality reduction and grouped convolution on the input feature map to obtain the intermediate feature map. The formula is as follows:

[0088]

[0089] in, This is an intermediate feature map; This is a grouped pointwise convolutional layer used for channel dimensionality reduction and feature recombination, reducing computational cost while preserving key channel information.

[0090] 2) Local feature extraction is performed on the intermediate feature map to obtain local detail features. The formula is as follows:

[0091]

[0092] in, For local detail features; It is a depthwise separable convolution used to capture spatial neighborhood relationships.

[0093] 3) Channel selection and nonlinear mapping are performed on local detailed features to obtain a feature result that fuses linear and nonlinear outputs, defined as a partial KAN, such as... Figure 4 As shown, the formula is expressed as:

[0094]

[0095] in, To integrate the characteristic results of linear and nonlinear outputs; Select a mask for the channel, only for Nonlinear mapping is applied to the channels to reduce computational costs; This is an element-wise multiplication operation; This provides learnable nonlinear transformation capabilities for KAN kernel function mapping. When When, the corresponding channel undergoes a KAN nonlinear transformation; when At that time, the channel remains linear.

[0096] Channel restoration and residual connection

[0097]

[0098] in, Output features for the module; This is a point-by-point convolutional layer used to restore the number of channels and fuse features; "This is a residual connection to ensure gradient stability and feature consistency."

[0099] Wherein, the KAN kernel function is:

[0100]

[0101] in, These are the feature values ​​of the input channel; For the first Segment spline weights; The segment node position; The number of segments; Linear rectifier functions enable mappings to exhibit nonlinearity and sparsity.

[0102] A nonlinear activation structure based on spline kernels is adopted to replace the traditional activation function. The activation function is decomposed into a weighted sum of multiple adjustable linear segments, and the weights of each segment are automatically learned during training. With nodes This results in an adaptive nonlinear response curve.

[0103] When the input features are small (smooth regions), The response is close to linear, which helps maintain smoothness at low frequencies; when the input features change drastically (texture or edge regions), It outputs a stronger nonlinear response to enhance local details and contrast.

[0104] 4. Based on the predicted noise term With structural residuals Based on the uncertainty diagram Dynamically fuse the two outputs to obtain the reconstructed image:

[0105]

[0106] in, For the first Step-by-step reconstruction of the image; The adaptive scheduling weights for residual branches; This represents the noise ratio of the current step. To predict noise; To predict residuals.

[0107] 5. Map the reconstructed image from the spatial domain to the frequency domain to obtain a frequency domain representation; divide the frequency domain representation into frequencies and assign learnable weights to each frequency band; construct frequency domain constraint terms based on the weighted results.

[0108] 1) The reconstructed image is transformed from the spatial domain to the frequency domain, expressed by the formula:

[0109]

[0110] in, This represents the channel average pooling characteristic. This represents the channel max pooling feature; For splicing operations; Linear fusion layer; : Sigmoid normalization; X is the reconstructed image.

[0111] 2) The loss weights for each frequency band are dynamically adjusted using an attention mask, as expressed by the formula:

[0112]

[0113]

[0114] in: All are learnable parameters. It is the weight matrix of the convolution. These are the corresponding bias terms; different Corresponding to different frequency bands (low, medium, and high frequencies); This is a mapping function. It is a learnable spatial mask that represents which regions are more important in the frequency domain, used to dynamically adjust the loss weights for each frequency band. Through convolution, the model automatically learns the regions of interest in each frequency band. Within the frequency domain constraints, Instead of directly calculating the loss, it is converted into frequency band weights through a mapping function.

[0115] when A larger value indicates that the region contains rich structure or texture, thus increasing the corresponding high-frequency weights. ;

[0116] when A smaller value indicates a smoother region, thus increasing the weighting of low-frequency frequencies. Suppress high-frequency noise.

[0117] 6. Model Training:

[0118] The first stage of training describes the process of obtaining the reconstructed image; the loss function for the first stage is:

[0119]

[0120] in, For pixel reconstruction items, For structure preservation terms; For frequency domain constraints;

[0121] The second stage trains the overall low-light image enhancement model, and the loss function for the second stage is:

[0122]

[0123] in: Image reconstruction loss; For frequency domain constraints; Regularization is guided by uncertainty.

[0124] 1) Uncertainty-driven regularization:

[0125]

[0126] Among them, if High values ​​correspond to regions where the model is "unreliable," where the regularization term has a large weight, forcing the predicted noise to match the actual noise; if In regions where the low-correspondence model has stabilized, reduce constraints to prevent overfitting.

[0127] 2) The image reconstruction loss is:

[0128]

[0129] in, The reconstructed image obtained from model inversion; The original low-light input image; For the first The image after adding noise; Output for residual branches; This is a pixel-level uncertainty map for the corresponding time step; The variance of the uncertainty plot is used to measure the dispersion of prediction confidence. These are the respective balancing weights, used to coordinate pixel consistency, residual consistency, and uncertainty stability; This is a pixel consistency term to ensure that the brightness of the inverted image is consistent with that of the real image; As a residual consistency term, it constrains the model's predicted residuals to be close to the actual residuals, thereby improving the ability to reconstruct structural details. As an uncertainty stabilizing term, it suppresses the variance of the pixel-level uncertainty map and ensures the stability of the sampling process.

[0130] 3) The frequency domain constraint loss is:

[0131]

[0132] in, Low / medium / high frequency bands; For the reason Dynamically adjusted bandwidth weights; These are phase balance parameters; The amplitude spectrum contains brightness and contrast information; The phase spectrum contains structural and edge information. The diffusion model is represented in the first... -1 step generated reconstructed image; Indicates a reference to the original image; 、 : respectively the first The amplitude spectrum of an image in each frequency band (low frequency / mid frequency / high frequency) reflects the brightness and contrast distribution information; 、 : respectively the first The phase spectrum of an image in each frequency band reflects structural, edge, and texture information; Spatial attention mask Through linear mapping function The generated frequency band weighting coefficients are used to dynamically adjust the contribution of each frequency band to the loss; This is the phase balance parameter, used to control the relative weights of the amplitude and phase losses. express Norm, used to measure the absolute distance of spectral differences.

[0133] 4) Structure preservation items, including:

[0134]

[0135] in, This is a structure-preserving term used to maintain the consistency of the edges and structure of the original image during the enhancement process; This represents the gradient map of the reconstructed image during the diffusion inversion stage. This represents the gradient map of the input low-light image; :for Norm, used to measure gradient difference.

[0136] Example

[0137] A visual perception and control method for wheeled-legged robot inspection has been applied to a quadruped robot for inspecting port conveyor belts. This robot is a biomimetic autonomous mobile platform with adaptive walking capabilities in complex terrain and multi-angle field-of-view stabilization control. Its path planning enables comprehensive inspection and multi-angle framing of key areas such as conveyor belt channels, idlers, and return belts. Standard mounting positions and power and network interfaces are reserved on the top / side of the platform. A customized vibration-damping bracket reliably connects to the gimbal-camera assembly, ensuring assembly accuracy, center of gravity stability, and vibration resistance, thus ensuring stable image acquisition under acceleration, turning, and ramp conditions.

[0138] A quadruped inspection robot: equipped with autonomous navigation, path planning, and stable motion control; an industrial high-definition pan-tilt camera: mounted on top of the robot, capable of real-time acquisition of the conveyor belt's operation, supporting automatic exposure and infrared illumination; an edge computing module (Jetson Orin NX): an embedded GPU computing unit for real-time model inference; a wireless communication module: enabling image feedback and control command interaction with the central monitoring system; and a main control terminal and remote monitoring platform: used for image storage, enhanced result display, and fault alarm management. The camera is mounted on top of the robot's pan-tilt unit with a stable bracket, covering the conveyor belt's operating area. The edge computing module and camera are connected via a high-speed interface (USB 3.0 / Ethernet) to ensure real-time input of video frames to the model inference module.

[0139] High-definition industrial cameras are used to capture inspection images under complex working conditions such as low light, dust, and shielding.

[0140] High-quality enhancement of low-light images using a low-light image enhancement model.

[0141] Inspection scene characteristics: open or semi-enclosed space with dust, shadows and uneven lighting;

[0142] Light intensity: Approximately 5–20 lux at night;

[0143] Inspection targets: structural abnormalities such as belt surface damage, belt misalignment, material spillage, and material accumulation;

[0144] Image type: RGB video frame (1280×720 / 1920×1080).

[0145] After being processed by the enhancement model, the brightness of the low-light image is increased by about 1.8–2.2 times, and the structural details are clearly discernible, effectively supporting the accurate identification of subsequent defect detection algorithms.

[0146] Data Acquisition and Preprocessing:

[0147] The system inspects and collects raw low-light video frames along a set path, and then uses normalization and white balance correction to generate standardized input.

[0148] Training the diffusion model:

[0149] The diffusion model was trained using LOL-v2 and a self-built belt conveyor dataset. The number of diffusion steps was set to 1000, and a linear-cosine hybrid strategy was used for noise scheduling. Uncertainty-guided regularization was introduced during training, and the reconstruction loss was weighted by pixel uncertainty maps, enabling the model to focus on learning in extremely dark areas.

[0150] Results and evaluation metrics:

[0151] The enhanced image is then denormalized and color restored before being output. The system automatically calculates metrics such as PSNR, SSIM, and LPIPS. The comparison results with existing methods are shown in Table 1.

[0152] The average PSNR of this method is 28.6 dB (an improvement of 9.1 dB over the original image); SSIM is improved from 0.63 to 0.91; and LPIPS is reduced by 42%.

[0153] Table 1

[0154]

[0155] Therefore, the enhanced image significantly improves brightness and detail.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A visual perception and control method applicable to inspection of wheeled robots, characterized in that, include: Acquire raw, low-light images of industrial equipment; The original low-light image is input into the low-light image enhancement model for image enhancement. The low-light image enhancement model processing procedure includes: The image is input into the U-net network to predict noise; The original image is denoised based on the predicted noise to obtain a reconstructed image as the enhanced image. The reconstructed image is mapped from the spatial domain to the frequency domain to obtain a frequency domain representation; the frequency domain representation is divided into frequencies, and learnable weights are assigned to each frequency band; a frequency domain constraint term is constructed based on the weighting result. The U-net network includes: The convolutional residual blocks in the feature extraction backbone between the encoder and decoder of the U-net network are replaced by a lightweight kernel activation function module. The lightweight kernel activation function module includes: Channel selection and nonlinear mapping are performed on local detail features to obtain a feature result that fuses linear and nonlinear outputs, expressed by the formula: in, To integrate the characteristic results of linear and nonlinear outputs; Select a mask for the channel, only for Nonlinear mapping is applied to the channels to reduce computational costs; This is an element-wise multiplication operation; It provides learnable nonlinear transformation capabilities for KAN kernel function mapping; when When, the corresponding channel undergoes a KAN nonlinear transformation; when At that time, the channel remains linear; Wherein, the KAN kernel function is: in, These are the feature values ​​of the input channel; For the first Segment spline weights; The segment node position; The number of segments; Linear rectifier functions enable mappings to exhibit nonlinearity and sparsity; The frequency domain image is divided into frequency bands and learnable weights are assigned to each band, including: The loss weights for each frequency band are dynamically adjusted using an attention mask, as expressed by the formula: in, This represents the channel average pooling characteristic. This represents the channel max pooling feature; All are learnable parameters. It is the weight matrix of the convolution. These are the corresponding bias terms; different Corresponding to different frequency bands; For mapping functions; It is a learnable spatial mask used to dynamically adjust the loss weights for each frequency band. In frequency domain constraints, It is not directly included in the loss calculation.

2. The method according to claim 1, characterized in that, The reconstructed image is obtained by denoising the original image based on the predicted noise. include: in, For the first Step-by-step reconstruction of the image; The adaptive scheduling weights for residual branches; This represents the noise ratio of the current step. To predict noise; To predict residuals.

3. The method according to claim 1, characterized in that, The training process of the low-light image enhancement model includes: The process of obtaining the reconstructed image described in the first stage of training; The loss function for the first stage is: in, For pixel reconstruction items, For structure preservation terms; For frequency domain constraints; The second stage trains the overall low-light image enhancement model, and the loss function for the second stage is: in: Image reconstruction loss; For frequency domain constraints; Regularization is guided by uncertainty.

4. The method according to claim 3, characterized in that, The frequency domain constraint: in, Low / medium / high frequency bands; For the reason Dynamically adjusted bandwidth weights; These are phase balance parameters; The amplitude spectrum includes brightness and contrast information; The phase spectrum contains structural and edge information; The diffusion model is represented in the first... -1 step to generate the reconstructed image; Indicates a reference to the original image; 、 : respectively the first The amplitude spectrum of an image in each frequency band reflects the brightness and contrast distribution information. 、 : respectively the first The phase spectrum of an image in each frequency band reflects structural, edge, and texture information; Spatial attention mask Through linear mapping function The generated frequency band weighting coefficients are used to dynamically adjust the contribution of each frequency band to the loss; This is the phase balance parameter, used to control the relative weights of the amplitude and phase losses. express Norm, used to measure the absolute distance of spectral differences.

5. The method according to claim 3, characterized in that, The uncertainty guides regularization: Among them, if In regions where high noise levels lead to model unreliability, the regularization term has a large weight, forcing the predicted noise to match the actual noise; if... Reduce constraints in regions where the model has already stabilized to prevent overfitting.

6. The method according to claim 3, characterized in that, The image reconstruction loss: in, The reconstructed image obtained from model inversion; The original low-light input image; For the first The image after adding noise; Output for residual branches; This is a pixel-level uncertainty map for the corresponding time step; The variance of the uncertainty plot is used to measure the dispersion of prediction confidence. These are the respective balancing weights, used to coordinate pixel consistency, residual consistency, and uncertainty stability; This is a pixel consistency term to ensure that the brightness of the inverted image is consistent with that of the real image; As a residual consistency term, it constrains the model's predicted residuals to be close to the actual residuals, thereby improving the ability to reconstruct structural details. As an uncertainty stabilizing term, it suppresses the variance of the pixel-level uncertainty map and ensures the stability of the sampling process.

7. The method according to claim 3, characterized in that, The structure preservation items include: in, This is a structure-preserving term used to maintain the consistency of the edges and structure of the original image during the enhancement process; This represents the gradient map of the reconstructed image during the diffusion inversion stage. This represents the gradient map of the input low-light image; :for Norm, used to measure gradient difference.

Images