Low-light abnormal railway state detection method based on multi-light noise field diffusion

Abstract

The present invention discloses a method for detecting abnormal railway conditions under low illumination based on diffusion of multiple illumination noise fields. For low illumination railway images, the present invention adopts histogram equalization to obtain a preliminary illumination correction image; adopts a Gaussian smoothing strategy to estimate the standard deviation of image noise to obtain a low noise correction image; adopts multi-gamma correction to generate a correction image under multiple noise distributions; uses the YOLO convolution feature of the illumination correction image to estimate the multiple noise distribution field, and adopts a diffusion model to suppress illumination noise in the feature; and uses a multi-branch YOLO detector to perform abnormal railway condition detection under multiple noise distributions. The present invention estimates multiple illumination noise fields and uses a diffusion model to suppress illumination noise, which can improve the accuracy of abnormal railway condition detection.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and information technology, and in particular to a low-light abnormal railway state detection method based on multi-light noise field diffusion. Background Art

[0002] With the construction of my country's "M"-shaped high-speed rail network, an eight-vertical and eight-horizontal structure has gradually been formed. The operation safety of trains has become a noteworthy issue. Researching the anomaly detection method of railway tracks has great value and significance. This can not only ensure the safety of railway operation, but also promote the development and progress of science and technology, and even more importantly, ensure the safety of people's lives and property.

[0003] Chinese Patent Publication No. CN113640308A "A Track Anomaly Monitoring System Based on Machine Vision". This system mainly uses machine vision to establish a neural network to identify faults. The specific steps of neural network training and fault classification output include: 1. Extracting a geometric feature data set and randomly selecting a sample point as the initial cluster center. 2. Calculating the shortest distance between each sample point and the existing cluster center, and using the roulette method to select the next new cluster center. 3. Calculating the center distance of each sample point to the new cluster center, and dividing the sample point into the category corresponding to the nearest cluster center. 4. Loop through the steps of recalculating the category center and dividing the sample points until the cluster center no longer changes and the minimum value is obtained. 5. Output the classification result, that is, the fault result, and perform targeted maintenance based on the result. This method not only improves the accuracy and efficiency of track monitoring, but also reduces the cost of track monitoring and maintenance, thereby ensuring track safety.

[0004] Chinese Patent Publication No. CN118072233A "A single-machine track foreign object detection method based on computer vision and radar fusion". This method mainly uses the YOLOv7 visual algorithm to perform target and anomaly detection to identify the types of foreign objects invading the track around the clock and automatically measure the distance; at the same time, when foreign objects are detected, timely feedback is given to improve work efficiency. First, data is collected and labeled, and after pre-processing, it is timely classified into data sets and training sets; then the classified training set is imported into the yolov7 algorithm, and the yolov7 algorithm is used to establish the yolov7 network structure detection system; then the algorithm parameters are tuned to achieve the best visual detection model state, and finally the algorithm's hardware and software are deployed; if the track is abnormal, an alarm feedback is given in the software.

[0005] Chinese Patent Publication No. CN116539723A "A Railway Track Damage Detection Method and System". This method utilizes ultrasonic flaw detection and visual images for abnormality detection. First, based on the ultrasonic information, a preset neural network model is used to obtain the railway rail head damage detection result, and the neural network is used to reduce the error, thereby avoiding the problem of poor amplitude restoration effect in the traditional ultrasonic detection method process; based on the rail surface image information, a preset machine vision algorithm is used to obtain the railway track surface damage detection result; then, based on the fuzzy model, the railway rail head damage detection result and the railway track surface damage detection result are evaluated for damage, and the final railway track damage detection result is obtained, which solves the interference of the metal layer on the surface of the rail on the detection result and improves the accuracy of the railway track damage evaluation.

[0006] In the field of rail anomaly detection, many methods have been proposed to detect this problem. However, most methods are limited to ideal conditions and fail to fully consider environmental factors, which have a certain impact on anomaly detection. For example, other environmental factors such as light. However, most current methods usually only use ideal conditions for processing and judgment, ignoring the judgment of the environment on anomaly detection. Summary of the invention

[0007] The purpose of the present invention is to make up for the defects of the existing technology. By exploring the detection of abnormal rail conditions under low light conditions, a low-light abnormal railway state detection method based on the diffusion of multi-light noise fields is provided. The present invention estimates the multi-light noise field and uses a diffusion model to suppress light noise, which can improve the accuracy of abnormal railway state detection and has practical application value.

[0008] The present invention is achieved through the following technical solutions:

[0009] A method for detecting abnormal railway status under low illumination based on multi-illumination noise field diffusion specifically comprises the following steps:

[0010] S1: Construct railway abnormal image dataset;

[0011] S2: Extract multi-illumination correction features;

[0012] S3: Extract single illumination diffusion features;

[0013] S4: Construct a railway anomaly detection model based on multi-illumination diffusion;

[0014] S5: Detect railway anomalies;

[0015] The extraction of multi-illumination correction features described in step S2 specifically includes the following steps:

[0016] S2-1: For railway anomaly image xi , use histogram equalization to obtain a preliminary illumination correction image x i represents the i-th railway abnormal image, 1≤i≤n i , n i Indicates the number of abnormal railway images, E stands for equalized, and is placed in the superscript to distinguish - x before equalization i , after equalization

[0017] S2-2: Acquiring a low-noise corrected image S stands for Gaussian smoothing, which is placed in the superscript to distinguish it from the Gaussian smoothing. After Gaussian smoothing

[0018] S2-3: Obtain multi-illumination correction image x i,j , 1≤j≤n j , n j Indicates the number of Gamma values;

[0019] S2-4: For multi-illumination correction image x i,j , perform the first layer of YOLO convolution to extract multi-light correction features

[0020] The extraction of single illumination diffusion features described in step S3 specifically includes the following steps:

[0021] S3-1: For multi-illumination correction features Input FC layer for encoding to obtain illumination prediction features

[0022] S3-2: Calculate lighting prediction features Sampling starting characteristics after diffusion M steps

[0023] S3-3: Extracting light diffusion features

[0024] The step S1 of constructing a railway abnormal image dataset to obtain a railway abnormal image dataset specifically includes the following steps:

[0025] S1-1: During the UAV railway inspection, a depth camera located in the center of the UAV is used to obtain railway anomaly videos;

[0026] S1-2: Extracting an abnormal frame set X from the railway abnormal video, X = {x i}, x i represents the i-th railway abnormal image, 1≤i≤n i , n iIndicates the number of railway anomaly images;

[0027] S1-3: Label the railway anomaly category label Y, Y = {y i},y i ∈{1,2,3,4}, label 1 indicates rail breakage, label 2 indicates bolt loss, label 3 indicates sleeper damage, and label 4 indicates foreign matter intrusion.

[0028] Step S2-2 of obtaining a low noise corrected image The specific steps include:

[0029] S2-2-1: Define the Gaussian kernel function G σ (x,y), where σ is the standard deviation of the Gaussian kernel;

[0030] S2-2-2: Preliminary illumination correction image Apply the Gaussian kernel function G σ (x,y) is Gaussian smoothed to obtain a low-noise corrected image

[0031] Step S2-3 of obtaining the multi-illumination correction image x i,j , 1≤j≤n j , n j Indicates the number of Gamma values, including the following steps:

[0032] S2-3-1: Define a set of Gamma values ​​γ = {γ j}, where γ j ≥0,1≤j≤n j , γ j represents the jth Gamma value, n j Indicates the number of Gamma values;

[0033] S2-3-2: For low noise correction images and Gamma valueγ j , apply multiple gamma correction operations Get the multi-illumination corrected image x of the i-th low-noise corrected image i,j .

[0034] Calculating illumination prediction features as described in step S3-2 Sampling starting characteristics after diffusion M steps The specific steps include:

[0035] S3-2-1: Initialize the sampling process with a sampling number N=100;

[0036] S3-2-2: Initialize step size h=5;

[0037] S3-2-3: Calculate the maximum number of times the diffusion sample is added with noise M = N*h;

[0038] S3-2-4: Initialize the lower limit of noise adjustment parameter β min =0.1, upper limit of noise adjustment parameter β max =20;

[0039] S3-2-5: Calculate the noise adjustment parameter β M =β min +M(β max -β min );

[0040] S3-2-6: Initialization noise Where w represents Gaussian noise with a mean of 0 and a variance of 1, and I represents a matrix whose elements are all 1;

[0041] S3-2-7: Calculate illumination prediction features Sampling starting characteristics after diffusion M steps

[0042]

[0043] Extracting the light diffusion feature described in step S3-3 The specific steps include:

[0044] S3-3-1: Calculate illumination prediction features Gaussian noise features after diffusion s steps

[0045] S3-3-2: Extracting illumination-dependent features

[0046] S3-3-3: Extracting illumination-dependent features

[0047] S3-3-4: Calculate lighting characteristics

[0048] S3-3-5: Calculate the modified mean eigenvector

[0049] S3-3-6: Calculate the eigenvector Features after denoising iteration

[0050] S3-3-7: When t n ≠0, n=n+1, and repeat steps S3-3-1 to S3-3-6. When t n = 0, the output is a single illumination diffusion feature

[0051] Calculating illumination prediction features as described in step S3-3-1 Gaussian noise features after diffusion s steps The specific steps include:

[0052] S3-3-1-1: Initialize step size h=5;

[0053] S3-3-1-2: Calculate the number of noise additions s = (Nn)*h, 0≤n≤N;

[0054] S3-3-1-3: Initialize the lower limit of noise adjustment parameter β min =0.1, upper limit of noise adjustment parameter β max =20;

[0055] S3-3-1-4: Calculate the noise adjustment parameter β s =β min +s(β max -β min );

[0056] S3-3-1-5: When s≠0, initialize the noise Where w represents Gaussian noise with mean 0 and variance 1. When s = 0, the noise w = 0;

[0057] S3-3-1-6: Calculate illumination prediction features Gaussian noise features after diffusion s steps

[0058]

[0059] Extracting illumination-dependent features as described in step S3-3-2 The specific steps include:

[0060] S3-3-2-1: Predicting features from illumination Input FC layer as query feature Q;

[0061] S3-3-2-2: Gaussian noise feature Input FC layer as key feature K s ;

[0062] S3-3-2-3: Gaussian noise feature Input FC layer as value feature V s ;

[0063] S3-3-2-4: Extracting illumination-dependent features

[0064] d represents the Gaussian noise feature The dimension of , T represents the matrix transpose;

[0065] Extracting illumination-dependent features as described in step S3-3-3 The specific steps include:

[0066] S3-3-3-1: Initialize t n =(Nn)*h, 0≤n≤N, when n=0, t n The value is M;

[0067] S3-3-3-2: Predicting features from illumination Input FC layer as query feature Q;

[0068] S3-3-3-3: Gaussian noise feature Input FC layer as key feature

[0069] S3-3-3-4: Gaussian noise feature Input FC layer as value feature

[0070] S3-3-3-5: Extracting illumination-dependent features

[0071]

[0072] Calculating illumination features as described in step S3-3-4 The specific steps include:

[0073] S3-3-4-1: Initialize constant λ ca =0.5;

[0074] S3-3-4-2: Calculate lighting characteristics

[0075]

[0076] Step S3-3-5 calculates the corrected mean eigenvector The specific steps include:

[0077] S3-3-5-1: Calculate mean features using a multi-residual layer network

[0078] S3-3-5-2: Calculate lighting characteristics Gradient

[0079] S3-3-5-3: Calculate mean corrected eigenvector

[0080]

[0081] Calculation characteristics described in step S3-3-6 Features after denoising iteration The specific steps include:

[0082] S3-3-6-1: Initialize step size h=5;

[0083] S3-3-6-2: Initialization noise Where w represents Gaussian noise with mean 0 and variance 1;

[0084] S3-3-6-3: Calculate β according to step S3-3-1-4 s ;

[0085] S3-3-6-4: Calculation characteristics Features after denoising iteration

[0086]

[0087] The construction of the railway anomaly detection model based on multi-light diffusion described in step S4 specifically includes the following steps:

[0088] S4-1: Diffusion of light features Enter the DarkNet53 network and obtain features of three different sizes

[0089] S4-2: Input the prediction head of YOLOv3 and output the prediction category score, railway anomaly category label and prediction box location information;

[0090] S4-3: Calculate the total loss function L based on the prediction head output i ;

[0091] S4-4: Using the total loss function L i As the gradient, the back propagation algorithm is used to train and obtain the optimal parameter set of the model, which includes the convolution kernel weight, bias, and anchor box width and height.

[0092] The total loss function L calculated in step S4-3 i , specifically including the following steps:

[0093] S4-3-1: Calculate the loss function

[0094] express The loss function of

[0095] S4-3-2: Calculate the loss function

[0096]

[0097] S4-3-3: Calculate the loss function

[0098]

[0099] S4-3-4: Calculate the total loss function L i :

[0100]

[0101] The detection of railway anomalies described in step S5 specifically includes the following steps:

[0102] S5-1: input the railway video frame to be tested;

[0103] S5-2: Extracting multi-illumination correction features according to step S2

[0104] S5-3: According to step S3, extract single illumination diffusion features

[0105] S5-4: According to step S4, use the model parameters obtained by training in S4-4 to perform target detection;

[0106] S5-5: Detect the location of railway anomalies.

[0107] The advantages of the present invention are as follows: for low-light railway images, the present invention adopts histogram equalization to obtain a preliminary illumination correction image; adopts a Gaussian smoothing strategy to estimate the image noise standard deviation to obtain a low-noise correction image; adopts multi-Gamma correction to generate a correction image under multiple noise distributions; uses the YOLO convolution feature of the illumination correction image to estimate the multiple noise distribution field, and adopts a diffusion model to suppress the illumination noise in the feature; uses a multi-branch YOLO detector to perform abnormal railway state detection under multiple noise distributions. The present invention estimates multiple illumination noise fields, uses a diffusion model to suppress illumination noise, fully considers environmental factors, and can improve the accuracy of abnormal railway state detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0108] Figure 1 It is a flow chart of the low-light abnormal railway state detection method based on multi-light noise field diffusion of the present invention;

[0109] Figure 2 The railway abnormal image data set construction process of the present invention;

[0110] Figure 3 The extraction process of the multi-illumination correction feature of the present invention;

[0111] Figure 4 The single illumination diffusion feature extraction process of the present invention;

[0112] Figure 5 The training process of the railway anomaly detection model of the present invention;

[0113] Figure 6 This is the railway anomaly detection process of the present invention. DETAILED DESCRIPTION

[0114] In order to make the purpose, technical solution and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. The present invention is a low-light abnormal railway state detection method based on multi-light noise field diffusion. The specific process is as follows Figure 1 As shown, the implementation scheme of the present invention is divided into the following steps:

[0115] S1: Construct railway abnormal image dataset X, such as Figure 2 As shown;

[0116] S1-1: During the UAV railway inspection, a depth camera located in the center of the UAV is used to obtain railway anomaly videos;

[0117] S1-2: Extract the abnormal frame set X from the video, X = {x i}(x i represents the i-th railway abnormal image, 1≤i≤n i , n i Indicates the number of railway abnormal images)

[0118] S1-3: Label the railway anomaly category label Y, Y = {y i},y i ∈{1,2,3,4} (label 1 indicates rail breakage, label 2 indicates bolt loss, label 3 indicates sleeper damage, label 4 indicates foreign matter intrusion);

[0119] S2: Extracting multi-illumination correction features like Figure 3 As shown;

[0120] S2-1: For railway anomaly image x i , use histogram equalization to obtain a preliminary illumination correction image

[0121] S2-2: Acquiring a low-noise corrected image

[0122] S2-2-1: Define the Gaussian kernel function G σ(x,y), where σ is the standard deviation of the Gaussian kernel.

[0123] S2-2-2: Preliminary illumination correction image Apply the Gaussian kernel function G σ (x,y) is Gaussian smoothed to obtain a low-noise corrected image

[0124] S2-3: Obtain multi-illumination correction image x i,j , 1≤j≤n j , n j Indicates the number of Gamma values;

[0125] S2-3-1: Define a set of Gamma values ​​γ = {γ j}, where γ j ≥0,1≤j≤n j , γ j represents the jth Gamma value, n j Indicates the number of Gamma values;

[0126] S2-3-2: For low noise correction images and Gamma valueγ j , apply multiple gamma correction operations Get the multi-illumination corrected image x of the i-th low-noise corrected image i,j ;

[0127] S2-4: For multi-illumination correction image x i,j , perform the first layer of YOLO convolution to extract multi-light correction features

[0128] S3: Extracting single illumination diffusion features like Figure 4 As shown;

[0129] S3-1: For multi-illumination correction features Input FC layer for encoding to obtain illumination prediction features

[0130] S3-2: Calculate lighting prediction features Sampling starting characteristics after diffusion M steps

[0131] S3-2-1: Initialize the sampling process with a sampling number N=100;

[0132] S3-2-2: Initialize step size h=5;

[0133] S3-2-3: Calculate the maximum number of times the diffusion sample is added to the noise M = N*h

[0134] S3-2-4: Initialize the lower limit of noise adjustment parameter β min =0.1, upper limit of noise adjustment parameter β max =20;

[0135] S3-2-5: Calculate the noise adjustment parameter β M =β min +M(β max -β min );

[0136] S3-2-6: Initialization noise Where w represents Gaussian noise with mean 0 and variance 1;

[0137] S3-2-7: Calculate illumination prediction features Sampling starting characteristics after diffusion M steps

[0138]

[0139] S3-3: Extracting light diffusion features

[0140] S3-3-1: Calculate illumination prediction features Gaussian noise features after diffusion s steps

[0141] S3-3-1-1: Initialize step size h=5;

[0142] S3-3-1-2: Calculate the number of noise additions s = (Nn)*h, 0≤n≤N;

[0143] S3-3-1-3: Initialize the lower limit of noise adjustment parameter β min =0.1, upper limit of noise adjustment parameter β max =20;

[0144] S3-3-1-4: Calculate the noise adjustment parameter β s =β min +s(β max -β min );

[0145] S3-3-1-5: When s≠0, initialize the noise Where w represents Gaussian noise with mean 0 and variance 1. When s = 0, the noise w = 0;

[0146] S3-3-1-6: Calculate illumination prediction features Characteristics after diffusion s steps

[0147]

[0148] S3-3-2: Extracting illumination-dependent features

[0149] S3-3-2-1: Predicting features from illumination Input FC layer as query feature Q;

[0150] S3-3-2-2: Gaussian noise feature Input FC layer as key feature K s ;

[0151] S3-3-2-3: Gaussian noise feature Input FC layer as value feature V s ;

[0152] S3-3-2-4: Extracting illumination-dependent features

[0153]

[0154] S3-3-3: Extracting illumination-dependent features

[0155] S3-3-3-1: Initialize t n =(Nn)*h, 0≤n≤N, when n=0, t n The value is M;

[0156] S3-3-3-2: Predicting features from illumination Input FC layer as query feature Q;

[0157] S3-3-3-3: Gaussian noise feature Input FC layer as key feature

[0158] S3-3-3-4: Gaussian noise feature Input FC layer as value feature

[0159] S3-3-3-5: Extracting illumination-dependent features

[0160]

[0161] S3-3-4: Calculate lighting characteristics

[0162] S3-3-4-1: Initialize constant λ ca =0.5;

[0163] S3-3-4-2: Calculate lighting characteristics

[0164]

[0165] S3-3-5: Calculate the modified mean eigenvector

[0166] S3-3-5-1: Calculate mean features using a multi-residual layer network

[0167] S3-3-5-2: Calculate lighting characteristics Gradient

[0168] S3-3-5-3: Calculate mean corrected eigenvector

[0169]

[0170] S3-3-6: Calculation characteristics Features after denoising iteration

[0171] S3-3-6-1: Initialize step size h=5;

[0172] S3-3-6-2: Initialization noise Where w represents Gaussian noise with mean 0 and variance 1;

[0173] S3-3-6-3: Calculate β according to step S3-3-1-4 s ;

[0174] S3-3-6-4: Calculation characteristics Features after denoising iteration

[0175]

[0176] S3-3-7: When t n ≠0, n=n+1, and repeat steps S3-3-1 to S3-3-6. n = 0, the output is a single illumination diffusion feature

[0177] S4: Construct a railway anomaly detection model based on multi-light diffusion, such as Figure 5 As shown;

[0178] S4-1: Diffusion of light features Enter the DarkNet53 network and obtain features of three different sizes

[0179] S4-2: Input the prediction head of YOLOv3 and output the prediction category score, railway anomaly category label and prediction box location information;

[0180] S4-3: Calculate the total loss function L based on the prediction head output i ;

[0181] S4-3-1: Calculate the loss function

[0182]

[0183] S4-3-2: Calculate the loss function

[0184]

[0185] S4-3-3: Calculate the loss function

[0186]

[0187] S4-3-4: Calculate the total loss function L i :

[0188]

[0189] S4-4: Using the total loss function L i As the gradient, the back propagation algorithm is used to train and obtain the optimal parameter set of the model, which includes the convolution kernel weight, bias, and anchor box width and height.

[0190] S5: Detect railway anomalies, such as Figure 6 As shown;

[0191] S5-1: input the railway video frame to be tested;

[0192] S5-2: Extracting multi-illumination correction features according to step S2

[0193] S5-3: According to step S3, extract single illumination diffusion features

[0194] S5-4: According to step S4, use the model parameters obtained by training in S4-4 to perform target detection;

[0195] S5-5: Detect the location of railway anomalies.

Claims

1. A method for detecting abnormal railway conditions under low illumination based on multi-illumination noise field diffusion, characterized in that: The specific steps include: S1: Construct railway abnormal image dataset; S2: Extract multi-illumination correction features; S3: Extract single illumination diffusion features; S4: Construct a railway anomaly detection model based on multi-illumination diffusion; S5: Detect railway anomalies; The extraction of multi-illumination correction features described in step S2 specifically includes the following steps: S2-1: For railway anomaly image x i , use histogram equalization to obtain a preliminary illumination correction image x i represents the i-th railway abnormal image, 1≤i≤n i , n i Indicates the number of railway anomaly images; S2-2: Acquiring a low-noise corrected image S2-3: Obtain multi-illumination correction image x i,j , 1≤j≤n j , n j Indicates the number of Gamma values; S2-4: For multi-illumination correction image x i,j , perform the first layer of YOLO convolution to extract multi-light correction features The extraction of single illumination diffusion features described in step S3 specifically includes the following steps: S3-1: For multi-illumination correction features Input FC layer for encoding to obtain illumination prediction features S3-2: Calculate lighting prediction features Sampling starting characteristics after diffusion M steps S3-3: Extracting light diffusion features 2. The method for detecting abnormal railway status under low illumination based on multi-illumination noise field diffusion according to claim 1 is characterized in that: The step S1 of constructing a railway abnormal image dataset to obtain a railway abnormal image dataset specifically includes the following steps: S1-1: During the UAV railway inspection, a depth camera located in the center of the UAV is used to obtain railway anomaly videos; S1-2: Extracting an abnormal frame set X from the railway abnormal video, X = {x i }, x i represents the i-th railway abnormal image, 1≤i≤n i , n i Indicates the number of railway anomaly images; S1-3: Label the railway anomaly category label Y, Y = {y i },y i ∈{1,2,3,4}, label 1 indicates rail breakage, label 2 indicates bolt loss, label 3 indicates sleeper damage, and label 4 indicates foreign matter intrusion.

3. The method for detecting abnormal railway status under low illumination based on multi-illumination noise field diffusion according to claim 2 is characterized in that: Step S2-2 of obtaining a low noise corrected image The specific steps include: S2-2-1: Define the Gaussian kernel function G σ (x,y), where σ is the standard deviation of the Gaussian kernel; S2-2-2: Preliminary illumination correction image Apply the Gaussian kernel function G σ (x,y) is Gaussian smoothed to obtain a low-noise corrected image The multi-illumination correction image x is obtained in step S2-3 i,j , 1≤j≤n j , n j Indicates the number of Gamma values, including the following steps: S2-3-1: Define a set of Gamma values ​​γ = {γ j }, where γ j ≥0,1≤j≤n j , γ j represents the jth Gamma value, n j Indicates the number of Gamma values; S2-3-2: For low noise correction images and Gamma valueγ j , apply multiple gamma correction operations Get the multi-illumination corrected image x of the i-th low-noise corrected image i,j .

4. The method for detecting abnormal railway status under low illumination based on multi-illumination noise field diffusion according to claim 1, characterized in that: Calculating illumination prediction features as described in step S3-2 Sampling starting characteristics after diffusion M steps The specific steps include: S3-2-1: Initialize the sampling process with a sampling number N=100; S3-2-2: Initialize step size h=5; S3-2-3: Calculate the maximum number of times the diffusion sample is added with noise M = N*h; S3-2-4: Initialize the lower limit of noise adjustment parameter β min =0.1, upper limit of noise adjustment parameter β max =20; S3-2-5: Calculate the noise adjustment parameter β M =β min +M(β max -β min ); S3-2-6: Initialization noise Where w represents Gaussian noise with a mean of 0 and a variance of 1, and I represents a matrix whose elements are all 1. It means it follows normal distribution; S3-2-7: Calculate illumination prediction features Sampling starting characteristics after diffusion M steps Extracting the light diffusion feature described in step S3-3 The specific steps include: S3-3-1: Calculate illumination prediction features Gaussian noise features after diffusion s steps S3-3-2: Extracting illumination-dependent features S3-3-3: Extracting illumination-dependent features S3-3-4: Calculate lighting characteristics S3-3-5: Calculate the modified mean eigenvector S3-3-6: Calculate the eigenvector Features after denoising iteration S3-3-7: When t n ≠0, n=n+1, and repeat steps S3-3-1 to S3-3-6. n = 0, the output is a single illumination diffusion feature 5. The method for detecting abnormal railway status under low illumination based on multi-illumination noise field diffusion according to claim 4 is characterized in that: Calculating illumination prediction features as described in step S3-3-1 Gaussian noise features after diffusion s steps The specific steps include: S3-3-1-1: Initialize step size h=5; S3-3-1-2: Calculate the number of noise additions s = (Nn)*h, 0≤n≤N; S3-3-1-3: Initialize the lower limit of noise adjustment parameter β min =0.1, upper limit of noise adjustment parameter β max =20; S3-3-1-4: Calculate the noise adjustment parameter β s =β min +s(β max -β min ); S3-3-1-5: When s≠0, initialize the noise Where w represents Gaussian noise with mean 0 and variance 1. When s = 0, the noise w = 0; S3-3-1-6: Calculate illumination prediction features Gaussian noise features after diffusion s steps Extracting illumination-dependent features as described in step S3-3-2 The specific steps include: S3-3-2-1: Predicting features from illumination Input FC layer as query feature Q; S3-3-2-2: Gaussian noise feature Input FC layer as key feature K s ; S3-3-2-3: Gaussian noise feature Input FC layer as value feature V s ; S3-3-2-4: Extracting illumination-dependent features d represents the Gaussian noise feature The dimension of , T represents the matrix transpose.

6. The method for detecting abnormal railway status under low illumination based on multi-illumination noise field diffusion according to claim 5, characterized in that: Extracting illumination-dependent features as described in step S3-3-3 The specific steps include: S3-3-3-1: Initialize t n =(Nn)*h, 0≤n≤N, when n=0, t n The value is M; S3-3-3-2: Predicting features from illumination Input FC layer as query feature Q; S3-3-3-3: Gaussian noise feature Input FC layer as key feature S3-3-3-4: Gaussian noise feature Input FC layer as value feature S3-3-3-5: Extracting illumination-dependent features Calculating illumination features as described in step S3-3-4 The specific steps include: S3-3-4-1: Initialize constant λ ca =0.5; S3-3-4-2: Calculate lighting characteristics Step S3-3-5 calculates the corrected mean eigenvector The specific steps include: S3-3-5-1: Calculate mean features using a multi-residual layer network S3-3-5-2: Calculate lighting characteristics Gradient S3-3-5-3: Calculate mean corrected eigenvector 7. The method for detecting abnormal railway status under low illumination based on multi-illumination noise field diffusion according to claim 6 is characterized in that: Calculation characteristics described in step S3-3-6 Features after denoising iteration The specific steps include: S3-3-6-1: Initialize step size h=5; S3-3-6-2: Initialization noise Where w represents Gaussian noise with mean 0 and variance 1; S3-3-6-3: Calculate β according to step S3-3-1-4 s ; S3-3-6-4: Calculation characteristics Features after denoising iteration 8. The method for detecting abnormal railway status under low illumination based on multi-illumination noise field diffusion according to claim 7 is characterized in that: The construction of the railway anomaly detection model based on multi-light diffusion described in step S4 specifically includes the following steps: S4-1: Diffusion of light features Enter the DarkNet53 network and obtain features of three different sizes S4-2: Input the prediction head of YOLOv3 and output the prediction category score, railway anomaly category label and prediction box location information; S4-3: Calculate the total loss function L based on the prediction head output i ; S4-4: Using the total loss function L i As the gradient, the back propagation algorithm is used to train and obtain the optimal parameter set of the model, which includes the convolution kernel weight, bias, and anchor box width and height.

9. The method for detecting abnormal railway status under low illumination based on multi-illumination noise field diffusion according to claim 8, characterized in that: The total loss function L calculated in step S4-3 i , specifically including the following steps: S4-3-1: Calculate the loss function express The loss function of S4-3-2: Calculate the loss function express The loss function of S4-3-3: Calculate the loss function express The loss function of S4-3-4: Calculate the total loss function L i :

10. The method for detecting abnormal railway status under low illumination based on multi-illumination noise field diffusion according to claim 1, characterized in that: The detection of railway anomalies described in step S5 specifically includes the following steps: S5-1: input the railway video frame to be tested; S5-2: Extracting multi-illumination correction features according to step S2 S5-3: According to step S3, extract single illumination diffusion features S5-4: According to step S4, use the model parameters obtained by training in S4-4 to perform target detection; S5-5: Detect the location of railway anomalies.

Images