Deep learning-based low-illumination transformer substation road damage detection method

Abstract

The invention belongs to the technical field of transformer substation roads, and particularly relates to a low-illumination transformer substation road damage detection method based on deep learning, and the method comprises the following specific steps: S1, constructing an original transformer substation road damage image data set D1; s2, preprocessing the D1 to obtain a processed transformer substation road damage image data set D2, performing low-illumination degradation synthesis operation on each transformer substation road damage image p in the D2 to obtain a low-illumination version image p ', and forming a transformer substation road damage image two-tuple lt; p, p'gt; . According to the method, reflectivity feature extraction and reflectivity optimization learning are introduced, so that the detection precision of transformer substation road damage in a low-illumination environment is improved, interference caused by equipment shading shadow and illumination change is effectively reduced, and the problems that an existing method excessively depends on an image enhancement technology under the low-illumination condition and noise is easily introduced are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of substation roads, and in particular to a method for detecting damage to substation roads in low light conditions based on deep learning. Background Art

[0002] Substation roads are critical infrastructure that ensures smooth equipment transportation and maintenance operations, and they are typically narrow. A large number of substation equipment, such as transformers and disconnectors, are distributed on both sides of the roads. The roads between these devices are used not only for inspections by maintenance personnel, but also for the passage of intelligent inspection equipment such as drones and unmanned vehicles. With the increase in substation maintenance operations and the increasing frequency of equipment transportation, the load on substation roads continues to increase, resulting in increasingly severe damage to substation roads. However, due to factors such as shadows created by equipment obstruction and insufficient light at night, damage to substation roads in low-light environments is difficult to detect in a timely manner, increasing the difficulty of substation road maintenance. Therefore, how to efficiently and accurately detect damage to substation roads under low-light conditions has become a technical challenge that needs to be addressed.

[0003] Currently, substation road damage detection primarily relies on manual inspections and deep learning methods. While manual inspections can detect obvious substation road damage, they are inefficient and costly, and due to the time and effort required for manual inspections, comprehensive coverage is difficult. Under normal lighting conditions, deep learning methods can efficiently and accurately identify road damage. However, substation roads are typically narrow, and equipment on both sides easily cast shadows, resulting in low light conditions on some roads. Furthermore, insufficient light at night or during inclement weather further exacerbates this problem. As a result, deep learning models trained under normal lighting conditions significantly degrade in low-light environments. Existing deep learning methods for detecting substation road damage in low-light conditions are primarily categorized into two types: detection-by-enhancement and enhancement-for-detection. Enhancement-based detection methods enhance low-light images (e.g., by increasing brightness and contrast) to improve image quality and thus enhance object detection accuracy. However, in low-light conditions, enhancement-based detection methods can introduce noise or artifacts, which can negatively impact detection performance. Detection enhancement methods combine image enhancement with the object detection process, optimizing training to better match the image enhancement effect with the detection task. Although detection enhancement methods improve detection accuracy, they still rely on enhancement. Furthermore, in low-light environments, this method's enhancement effect is limited, resulting in insufficient feature representation and an inability to fully address low-light road detection. While both methods improve the accuracy of low-light road damage detection under certain conditions, they generally rely on image enhancement, which has limited effectiveness in low-light conditions.

[0004] Based on the above, the current method has the following shortcomings:

[0005] 1. Substation roads are often exposed to low light due to shadows cast by substation equipment. Existing image enhancement techniques have limited effectiveness in low-light conditions and are prone to introducing noise. Therefore, existing methods are unable to effectively meet the needs of substation road detection in low-light environments.

[0006] 2. The existing methods have insufficient performance in detecting substation road damage in low-light environments, and are prone to missed detection or false detection, making it difficult to meet the actual needs of safe and efficient operation and maintenance of substations.

[0007] Therefore, a deep learning-based low-light substation road damage detection method was invented. By optimizing the extraction of reflectivity features, the model's ability to capture target features in low-light environments is enhanced, while reducing the interference of lighting changes and equipment shadows, further improving the model's detection accuracy and robustness, thereby providing strong technical support for the intelligent maintenance of substation roads. Summary of the Invention

[0008] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0009] A method for detecting road damage in low-light substations based on deep learning, comprising the following specific steps:

[0010] S1: Construct the original substation road damage image dataset D1;

[0011] S2: Preprocess D1 to obtain the processed substation road damage image dataset D2, and perform low-light degradation synthesis operation on each substation road damage image p in D2 to obtain a low-light version image p', and form a substation road damage image binary<p,p'> , a paired substation road damaged image dataset D3 is constructed from all substation road damaged image pairs;

[0012] S3: Build a deep learning-based low-light substation road damage detection model. Input the paired substation road damage image dataset D3 into the low-light substation road damage detection model to identify the road damage type in the substation road damage image and automatically generate a positioning frame to determine the specific location and scope of the damage. Road damage types include cracks, potholes, warping, and settlement.

[0013] S4: Before training the low-light substation road damage detection model, you first need to set hyperparameters. After completing the initial configuration, the paired substation road damage image dataset is divided into a training set and a validation set to ensure the independence and representativeness of each dataset.

[0014] S5: After model training is completed, the trained low-light substation road damage detection model is used to identify and analyze the current substation road damage data. The final output includes the specific substation road damage type and location identification results, providing substation inspection personnel with accurate early warning information and subsequent processing suggestions.

[0015] As a preferred solution of the method for detecting road damage in low-light substations based on deep learning described in the present invention, the specific steps of S1 are as follows:

[0016] S11: Collect images of damaged substation roads under normal lighting to form an original substation road damaged image dataset D1;

[0017] S12: Manually label D1 to provide labels for subsequent model training. During the labeling process, use the labeling tool to select the road damage area and add corresponding labels to each damaged area to mark the road damage type in the area.

[0018] As a preferred solution of the method for detecting road damage in a low-light substation based on deep learning described in the present invention, the specific steps of S2 are as follows:

[0019] S21: Crop all substation road damage images in D1 to a uniform size;

[0020] S22: Expand D1 by using data augmentation operations such as rotation, flipping, and noise addition to generate a substation road damage image dataset D2 containing more substation road damage images;

[0021] S23: Process each substation road damage image p in D2 by using brightness adjustment, contrast adjustment, color saturation reduction, blurring, and color shifting operations to generate a corresponding low-light version image p';

[0022] S24: Combine p and p' to form a substation road damage image binary<p,p'> , a paired substation road damaged image dataset D3 is constructed by all the above substation road damaged image binary pairs.

[0023] As a preferred solution of the method for detecting road damage in a low-light substation based on deep learning described in the present invention, the specific steps of S3 are as follows:

[0024] S31: For any substation road damage image pair<p,p'> ∈D3, in the first branch, p is input into the first Conv module to generate the first normal light road damage feature map X1, and p' is input into the first Conv module to generate the first low light road damage feature map X'1; in the second branch, p is input into the decomposition-based reflectivity and illumination decomposition module DRD to generate the first normal light road damage reflectivity feature map R1, and p' is input into the DRD module to generate the first low light road damage reflectivity feature map R'1;

[0025] S32:<X1,X'1> Input into the second Conv module to generate the second road damage feature Figure 2 Tuple<X2,X'2> , where X1 is input into the second Conv module to obtain the second normal light road damage feature map X2, and X'1 is input into the second Conv module to obtain the second low light road damage feature map X'2. Since the structure of the second Conv module is exactly the same as that of the first Conv module, the continuity and consistency of feature extraction can be ensured. On this basis,<X2,X′2> Passed to the CELAN module to generate a higher-level feature map, namely the third road damage feature Figure 2 Tuple<X3,X′3> , wherein X2 is input into the CELAN module to obtain the third normal-light road damage feature map X3, and X′2 is input into the CELAN module to obtain the third low-light road damage feature map X′3;

[0026] S33:<X2,X′2> Input into the DLARK module to generate the second road damage reflectivity feature Figure 2 Tuple<R2,R′2> , wherein X2 is input into the DLARK module to obtain the second normal light road damage reflectivity feature map R2, and X'1 is input into the DLARK module to obtain the second low light road damage reflectivity feature map R'2;

[0027] S34: In order to further distinguish the reflectance and illumination components in the image and maintain the consistency between illumination and reflectance,<R1,R'1> and<R2,R'2> Input into the illumination and reflection compensation module IRC, and generate the third road damage reflectivity feature by introducing the process of interaction, recombination and consistency constraints Figure 2 Tuple<R3,R'3> , wherein R1 and R2 are input into the IRC module to obtain a third normal light road damage reflectivity characteristic map R3, and R'1 and R'2 are input into the IRC module to obtain a third low light road damage reflectivity characteristic map R'3;

[0028] S35: Yes<R3,R′3> and<X3,X'3> Perform the Concat operation to fuse the reflectivity feature with the high-level semantic feature to obtain the fourth road damage feature. Figure 2 Tuple<X4,X'4> , wherein R3 and X3 are concat-operated to obtain the fourth normal-light road damage feature map X4, and R'3 and X'3 are concat-operated to obtain the fourth low-light road damage feature map X'4;

[0029] S36:<X4,X'4> The input is fed into the target prediction head to generate a tensor containing the prediction information. Each row of the tensor corresponds to a detection result, including the predicted bounding box coordinates, damage category label, and confidence score information.

[0030] As a preferred solution of the method for detecting road damage in a low-light substation based on deep learning described in the present invention, the overall execution process of the CELAN module in S32 is as follows:

[0031] S321: In the first branch,<T1,T'1> Input to the Conv2d layer with a convolution kernel size of 3×3 for convolution operation to generate the second intermediate road damage feature Figure 2 Tuple<T2,T'2> , where T1 is input into the Conv2d layer with a convolution kernel size of 3×3 to obtain the second intermediate normal light road damage feature map T2, and T'1 is input into the Conv2d layer with a convolution kernel size of 3×3 to obtain the second intermediate low light road damage feature map T'2, and then<T2,T'2> Input to the RepConv layer with a convolution kernel size of 3×3 for convolution operation to obtain the third intermediate road damage feature Figure 2 Tuple<T3,T'3> , where T2 is input to the RepConv layer with a convolution kernel size of 3×3 to obtain the third intermediate normal light road damage feature map T3, and T'2 is input to the RepConv layer with a convolution kernel size of 3×3 to obtain the third intermediate low light road damage feature map T'3, and then<T3,T'3> Input to the FreqConv layer with a convolution kernel size of 5×5 for convolution operation to generate the fourth intermediate road damage feature Figure 2 Tuple<T4,T'4> , where T3 is input into the FreqConv layer with a convolution kernel size of 5×5 to obtain the fourth middle normal light road damage feature map T4, and T'3 is input into the FreqConv layer with a convolution kernel size of 5×5 to obtain the fourth middle low light road damage feature map T'4, and then<T4,T'4> Input to the FreqConv layer with a convolution kernel size of 7×7 for convolution operation to generate the fifth intermediate road damage feature Figure 2Tuple<T5,T'5> , where T4 is input into the FreqConv layer with a convolution kernel size of 7×7 to obtain the fifth middle normal light road damage feature map T5, and T'4 is input into the FreqConv layer with a convolution kernel size of 7×7 to obtain the fifth middle low light road damage feature map T'5.<T5,T'5> Input to the Conv2d layer with a convolution kernel size of 3×3 for convolution operation to generate the sixth intermediate road damage feature Figure 2 Tuple<T6,T'6> , wherein T5 is input into the Conv2d layer with a convolution kernel size of 3×3 to obtain the sixth intermediate normal light road damage feature map T6, and T'5 is input into the Conv2d layer with a convolution kernel size of 3×3 to obtain the sixth intermediate low light road damage feature map T'6;

[0032] S322: In the second branch,<T1,T'1> Input into a continuous convolution structure for dilated convolution operation to obtain the seventh intermediate road damage feature Figure 2 Tuple<T7,T'7> , wherein T1 is input into the continuous convolution structure to obtain the seventh intermediate normal light road damage feature map T7, and T'1 is input into the continuous convolution structure to obtain the seventh intermediate low light road damage feature map T'7;

[0033] S323:<T2,T'2> 、<T3,T'3> 、<T4,T'4> 、<T5,T'5> 、<T6,T'6> and

[0034] <T7,T'7> Perform Concat operation to generate the eighth intermediate road damage feature Figure 2 Tuple<T8,T'8> , where T2, T3, T4, T5, T6 and T7 are concat-operated to obtain the eighth intermediate normal light road damage feature map T8, and T'2, T'3, T'4, T'5, T'6 and T'7 are concat-operated to obtain the eighth intermediate low light road damage feature map T'8, thereby achieving the effective fusion of multi-scale features and different receptive field information, further enhancing the comprehensive expression ability of the low light substation road damage detection model for detail features and global features in low light scenes;

[0035] S324:<T8,T'8> The input is sent to the Conv2d layer with a convolution kernel size of 5×5 for convolution operation and activated by the Sigmoid activation function to generate the ninth middle road damage feature. Figure 2Tuple<T9,T'9> , where T8 is input into the Conv2d layer with a convolution kernel size of 5×5 and activated by the Sigmoid activation function to obtain the ninth intermediate normal light road damage feature map T9, and T′8 is input into the Conv2d layer with a convolution kernel size of 5×5 and activated by the Sigmoid activation function to obtain the ninth intermediate low light road damage feature map T′9,<T9,T′9> As the final output of the CELAN module, it provides key input for subsequent feature extraction and prediction tasks.

[0036] As a preferred solution of the method for detecting road damage in low-light substations based on deep learning according to the present invention, the overall execution process of the DLARK module in S33 is as follows:

[0037] S331: In the first branch,<X2,X'2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 9×9 and a void rate of 1 and a BN layer for convolution and normalization operations to obtain the first intermediate reflectivity feature. Figure 2 Tuple <X 2_1 ,X' 2_1 >, where X2 is input into the convolution structure to obtain the first intermediate normal illumination reflectance feature map X 2_1 , input X'2 into the convolution structure to obtain the first intermediate low light reflectance feature map X' 2_1 ;

[0038] S332: In the second branch,<X2,X′2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 5×5 and a void rate of 1 and a BN layer for convolution and normalization operations to generate the second intermediate reflectivity feature. Figure 2 Tuple <X 2_2 ,X′ 2_2 >, where X2 is input into the convolution structure to obtain the second intermediate normal illumination reflectance feature map X 2_2 , input X′2 into the convolution structure to obtain the second intermediate low light reflectance feature map X′ 2_2 ;

[0039] S333: In the third branch,<X2,X′2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 5×5 and a void rate of 2 and a BN layer for convolution and normalization operations to obtain the third intermediate reflectivity feature. Figure 2 Tuple <X 2_3 ,X′ 2_3 >, where X2 is input into the convolution structure to obtain the third intermediate normal illumination reflectance feature map X 2_3 , input X′2 into the convolution structure to obtain the third intermediate low light reflectance feature map X′ 2_3;

[0040] S334: In the fourth branch,<X2,X′2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 2×2 and a void rate of 2 and a BN layer for convolution and normalization operations to generate the fourth intermediate reflectivity feature. Figure 2 Tuple <X 2_4 ,X′ 2_4 >, where X2 is input into the convolution structure to obtain the fourth intermediate normal illumination reflectance feature map X 2_4 , input X′2 into the convolution structure to obtain the fourth intermediate low light reflectance feature map X′ 2_4 ;

[0041] S335: In the fifth branch,<X2,X′2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 2×2 and a void rate of 4 and a BN layer for convolution and normalization operations to obtain the fifth intermediate reflectivity feature. Figure 2 Tuple <X 2_5 ,X' 2_5 >, where X2 is input into the convolution structure to obtain the fifth intermediate normal illumination reflectance feature map X 2_5 , input X'2 into the convolution structure to obtain the fifth intermediate low light reflectance feature map X' 2_5 ;

[0042] S336: Yes <X 2_1 ,X′ 2_1 >、 <X 2_2 ,X' 2_2 >、 <X 2_3 ,X' 2_3 >、 <X 2_4 ,X' 2_4 > and

[0043] <X 2_5 ,X' 2_5 >Perform the Concat operation to obtain the sixth intermediate reflectivity feature Figure 2 Tuple <X 2_6 ,X' 2_6 >, where X 2_1 、X 2_2 、X 2_3 、X 2_4 and X 2_5 Perform the Concat operation to obtain the sixth intermediate normal light reflectance feature map X 2_6 , X' 2_1 、X' 2_2 、X' 2_3 、X' 2_4 and X' 2_5Perform the Concat operation to obtain the sixth intermediate low light reflectance feature map X' 2_6 ,Will <X 2_6 ,X' 2_6 > Input to the AvgPool layer for pooling operation to obtain the seventh intermediate reflectivity feature Figure 2 Tuple <X 2_7 ,X' 2_7 >, where X 2_6 Input to the AvgPool layer to obtain the seventh intermediate normal light reflectance feature map X 2_7 , X′ 2_6 Input to the AvgPool layer to obtain the seventh intermediate low light reflectance feature map X' 2_7 ,Will <X 2_7 ,X' 2_7 > Input to the continuous convolution structure consisting of a Conv2d layer with a convolution kernel size of 1×1, a DWConv layer with a convolution kernel size of 2×2, a DWConv layer with a convolution kernel size of 5×5, and a Conv2d layer with a convolution kernel size of 2×2 for deep convolution operation to generate the eighth intermediate reflectivity feature Figure 2 Tuple <X 2_8 ,X' 2_8 >, where X 2_7 Input to the continuous convolution structure to obtain the eighth intermediate normal light reflectance feature map X 2_8 , X′ 2_7 Input to the continuous convolution structure to obtain the eighth intermediate low light reflectance feature map X′ 2_8 , in the right <X 2_6 ,X′ 2_6 >While performing the pooling operation, <X 2_6 ,X′ 2_6 > Input to the BN layer for normalization operation to generate the ninth intermediate reflectivity feature Figure 2 Tuple <X 2_9 ,X′ 2_9 >, where X 2_6 Input to the BN layer to obtain the ninth intermediate normal light reflectance feature map X 2_9 , X′ 2_6 Input to the BN layer to obtain the ninth intermediate low light reflectance feature map X′ 2_9 ;

[0044] S337: <X 2_9 ,X′ 2_9 > Input to the AvgPool layer for pooling operation to obtain the tenth intermediate reflectivity feature Figure 2 Tuple <X 2_10 ,X′ 2_10 >, where X 2_9Input to the AvgPool layer to obtain the tenth intermediate normal light reflectance feature map X 2_10 , X′ 2_9 Input to the AvgPool layer to obtain the tenth intermediate low light reflectance feature map X′ 2_10 , then <X 2_10 ,X′ 2_10 > Input into the structure consisting of a fully connected layer, a ReLU activation function and a fully connected layer sequence to perform feature transformation operations to obtain the eleventh intermediate reflectivity feature Figure 2 Tuple <X 2_11 ,X′ 2_11 >, where X 2_10 Input to the structure composed of the full connection layer, ReLU activation function and full connection layer sequence to obtain the eleventh intermediate normal light reflectance feature map X 2_11 , X′ 2_10 Input to the structure composed of a fully connected layer, a ReLU activation function and a fully connected layer in sequence to obtain the eleventh intermediate low light reflectance feature map X′ 2_11 ,Will <X 2_11 ,X′ 2_11 >Activate through the Sigmoid activation function to obtain the twelfth intermediate reflectivity feature Figure 2 Tuple <X 2_12 ,X′ 2_12 >, where X 2_11 The twelfth intermediate normal illumination reflectance feature map X is obtained by activating the Sigmoid function 2_12 , X′ 2_11 The twelfth intermediate low-light reflectance feature map X′ is obtained by activating the Sigmoid function 2_12 ,Will <X 2_9 ,X′ 2_9 >with <X 2_12 ,X′ 2_12 >Perform element-by-element multiplication to obtain the thirteenth intermediate reflectivity feature Figure 2 Tuple <X 2_13 ,X′ 2_13 >, where X 2_9 With X 2_12 Perform element-by-element multiplication to obtain the thirteenth intermediate normal light reflectance feature map X 2_13 , X′ 2_9 and X′ 2_12 Perform element-by-element multiplication to obtain the thirteenth intermediate low-light reflectance feature map X′ 2_13 ,Will <X 2_13 ,X′ 2_13 > Input into the structure consisting of a fully connected layer, a ReLU activation function and a fully connected layer sequence to perform feature transformation operation, and obtain the fourteenth intermediate reflectivity feature Figure 2Tuple <X 2_14 ,X′ 2_14 >, where X 2_13 Input to the structure composed of the full connection layer, ReLU activation function and full connection layer sequence to obtain the fourteenth intermediate normal light reflectance feature map X 2_14 , X′ 2_13 Input to the structure composed of a fully connected layer, a ReLU activation function and a fully connected layer in sequence to obtain the fourteenth intermediate low light reflectance feature map X′ 2_14 ;

[0045] S338: Yes <X 2_14 ,X′ 2_14 >After performing the DropPath operation, the fifteenth intermediate reflectivity feature is obtained Figure 2 Tuple <X 2_15 ,X′ 2_15 >, where X 2_14 Perform DropPath operation to obtain the fifteenth intermediate normal light reflectance feature map X 2_15 , X′ 2_14 Perform DropPath operation to obtain the fifteenth intermediate low light reflectance feature map X′ 2_15 ,Will <X 2_15 ,X′ 2_15 >with <X 2_8 ,X′ 2_8 >Perform the Concat operation to obtain the final second road damage reflectivity feature Figure 2 Tuple<R2,R′2> , where X 2_15 With X 2_8 Perform the Concat operation to obtain the second normal light road damage reflectivity feature map R2, and convert X′ 2_15 and X′ 2_8 Perform the Concat operation to obtain the second low-light road damage reflectivity feature map R′2.

[0046] As a preferred solution of the method for detecting road damage in a low-light substation based on deep learning described in the present invention, the specific steps of S4 are as follows:

[0047] S41: During the training process, the model extracts data from the training set according to the set batch size for forward calculation to generate prediction results. Subsequently, the model performs backpropagation through the overall loss function L2R-LOSS to optimize the parameters of the substation road damage detection model in low-light environments.

[0048] S42: After each round of training, the validation set data is used to evaluate the model to determine the generalization ability of the model;

[0049] S43: To prevent overfitting of the model during training, an early stopping mechanism is introduced during training. If the loss on the validation set does not improve significantly over multiple consecutive training rounds, training will be terminated early. This helps save computing resources and ensures the stability of the model in practical applications.

[0050] S44: When the training reaches the predetermined number of rounds and the validation set loss is stable or reaches an ideal level, the training process ends, and an optimized low-light substation road damage detection model is finally obtained. The model can accurately detect substation road damage under low-light conditions and maintain good performance in different environments;

[0051] S45: After training, the model will have appropriate parameter configuration and strong generalization ability, and can be applied to the new substation road damage detection task.

[0052] As a preferred solution of the method for detecting road damage in low-light substations based on deep learning described in the present invention, the loss function L2R-LOSS in S41 is composed of the reflectivity learning loss function L r , Reorganization consistency loss function L rc And the target detection loss function L target The weighted composition, its function expression is as follows:

[0053] L L2R-LOSS =λ g L r +λ f L target +λ q L rc

[0054] Among them, λ g ,λ f and λ q is the loss weight, the sum of the three is 1;

[0055] Among them, L r The function expression is as follows:

[0056]

[0057] Among them, R′1 is the first low-light road damage reflectivity feature map; R′2 is the first low-light road damage reflectivity feature map; N is the total number of feature points of the R′1 and R′2 feature maps, R′ 1_i and R′ 2_i Respectively represent the feature response value corresponding to the i-th position in the feature map; and denote the average characteristic response values of R′1 and R′2, respectively. and are the variances of the characteristic responses of R′1 and R′2, is the covariance of the characteristic response values between R'1 and R'2, and C1 and C2 are constants to avoid the denominator being zero;

[0058] Among them, L target The function expression is as follows:

[0059]

[0060] Among them, α and β represent the weights of position regression loss and classification loss respectively; x i and x' i Represents the center point of the real box and the predicted box of the i-th target respectively; p i and p' i They represent the true category label and predicted category probability of the i-th target, respectively, and Smooth(x) represents the error between the predicted value and the true value, that is,

[0061]

[0062] L rc The function expression is as follows:

[0063] L rc =||R1-R′2||1+||R′1-R2||1

[0064] Among them, R1 is the first normal light road damage reflectivity characteristic map; R'2 is the second low light road damage reflectivity characteristic map; R'1 is the first low light road damage reflectivity characteristic map; R2 is the second normal light road damage reflectivity characteristic map.

[0065] Compared with existing technologies:

[0066] By introducing reflectivity feature extraction and reflectivity optimization learning, the present invention improves the detection accuracy of substation road damage in low-light environments, effectively reduces the interference caused by equipment occlusion shadows and lighting changes, and avoids the problem of existing methods over-relying on image enhancement technology and easily introducing noise under low-light conditions; at the same time, combined with multi-scale feature extraction and attention mechanism, the model's adaptability and robustness to narrow substation roads and complex occlusion scenarios are enhanced, thereby improving the practicality and reliability of the low-light substation road damage detection model. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a schematic diagram of the process of the present invention;

[0068] Figure 2 This is a structural diagram of the low-light substation road damage detection model of the present invention;

[0069] Figure 3 This is a structural diagram of the CELAN module of the present invention;

[0070] Figure 4 This is the structural diagram of the DLARK module of the present invention. DETAILED DESCRIPTION

[0071] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0072] This invention provides a method for detecting road damage in low-light substations based on deep learning. Figures 1-4 , including the following specific steps:

[0073] S1: Construct the original substation road damage image dataset D1;

[0074] The specific steps of S1 are as follows:

[0075] S11: Collect images of damaged substation roads under normal lighting to form an original substation road damaged image dataset D1;

[0076] S12: Manually label D1 to provide labels for subsequent model training. During the labeling process, use the labeling tool to select the road damage area and add corresponding labels to each damaged area to mark the road damage type in the area;

[0077] Among them, S1 includes but is not limited to the following embodiments:

[0078] The dataset was processed based on a self-collected dataset of original substation road damage images. The dataset contains 2,000 images, including 800 images of road cracks, 500 images of potholes, 450 images of road lifts, and 250 images of road subsidence. Image labels can be divided into four categories based on damage type: cracks, potholes, lifts, and subsidence. The labeling tool LabelImg was used to manually annotate the image data and generate corresponding XML label files. The files contain the label names of the road damage and the corresponding target annotation box information.

[0079] The LabelImg tool is an open source image annotation tool, which is mainly used for image dataset annotation for tasks such as target detection and image segmentation.

[0080] S2: Preprocess D1 to obtain the processed substation road damage image dataset D2, and perform low-light degradation synthesis operation on each substation road damage image p in D2 to obtain a low-light version image p', and form a substation road damage image binary<p,p′> , a paired substation road damaged image dataset D3 is constructed from all substation road damaged image pairs;

[0081] The specific steps of S2 are as follows:

[0082] S21: Crop all substation road damage images in D1 to a uniform size;

[0083] S22: Expand D1 by using data augmentation operations such as rotation, flipping, and noise addition to generate a substation road damage image dataset D2 containing more substation road damage images;

[0084] S23: Process each substation road damage image p in D2 by using brightness adjustment, contrast adjustment, color saturation reduction, blurring, and color shifting operations to generate a corresponding low-light version image p′;

[0085] S24: Combine p and p′ into a substation road damage image binary<p,p′> , a paired substation road damaged image dataset D3 is constructed from all the above substation road damaged image pairs;

[0086] Among them, S2 includes but is not limited to the following embodiments:

[0087] The original substation road damage image dataset D1 contains a total of 2000 images. First, all images are uniformly cropped to 512×512 pixels. Then, a variety of data enhancement methods are used to expand the dataset, including rotation, flipping, and noise addition, which can be implemented in Python. Among them, the rotation operation is implemented by the cv2.getRotationMatrix2D() function; the image flipping is implemented by the cv2.flip() function; and the noise addition is implemented by the cv2.randPepper() function. After the above enhancement processing, D1 is expanded to the substation road damage image dataset D2 containing 12,000 images. On this basis, in order to simulate the road damage scene in low-light environment, each image in D2 is further The images were processed for low-light conditions. Specific operations included brightness adjustment, contrast adjustment, color desaturation, blurring, and color shifting. Brightness adjustment was performed using the ImageEnhance.Brightness() method; contrast adjustment was performed using ImageEnhance.Contrast(); color desaturation was achieved using ImageEnhance.Color(); blurring was performed using the ImageFilter.GaussianBlur() method; and color shifting was accomplished by modifying the pixel values of the image's RGB channels. Finally, the low-light versions of the images were paired with the original normal-light images to construct a paired substation road damage image dataset, D3, totaling 24,000 images.

[0088] The original substation road damage image dataset D1 is expanded, and the common operations used include the following:

[0089] (1) Rotation: Rotation simulates a change in viewing angle by rotating the image by a certain angle. You can use the rotate() method of the PIL library to randomly select an angle range (e.g., -30° to 30°) and rotate the image around the center point. This method can effectively enhance the model's ability to recognize images at different angles.

[0090] (2) Flip: The flip operation is achieved by mirroring the image along the horizontal or vertical axis. Horizontal flipping can be achieved using the PIL library's transpose(Image.FLIP_LEFT_RIGHT) method, while vertical flipping can be achieved using transpose(Image.FLIP_TOP_BOTTOM) or cv2.flip(). This operation can enhance the model's ability to recognize different directions.

[0091] (3) Noise Injection: The noise injection operation simulates the degradation of image quality by adding Gaussian noise or salt and pepper noise to the image. You can use numpy to generate a Gaussian noise matrix of the same size as the image and add it to the image pixel values, or use the cv2.randSalt() and cv2.randPepper() functions to generate salt and pepper noise. This operation helps improve the robustness of the model to low-quality images.

[0092] The low-light degradation synthesis technique is used to process the substation road damage image dataset D2. The common operations used include the following:

[0093] (1) Brightness Adjustment: Brightness adjustment simulates different lighting conditions by changing the brightness of an image. You can use the PIL.ImageEnhance.Brightness() method to adjust the brightness of an image by setting an enhancement factor. When the enhancement factor is greater than 1, the image becomes brighter, and when it is less than 1, the image becomes darker. This operation helps simulate the visual effects of nighttime or dim environments.

[0094] (2) Contrast Adjustment: Contrast adjustment simulates different ambient lighting conditions by enhancing or reducing the difference between light and dark in an image. Using the PIL.ImageEnhance.Contrast() method, you can set an enhancement factor to adjust the contrast. When the enhancement factor is greater than 1, the image contrast increases, and when it is less than 1, the contrast decreases. This operation helps simulate changes in the visibility of objects under different lighting conditions.

[0095] (3) Saturation Adjustment: Saturation adjustment simulates the visual effects of low light or different lighting conditions by changing the saturation of the image colors. You can use the PIL.ImageEnhance.Color() method to adjust the saturation of an image. Increasing saturation makes the colors more vivid, while decreasing saturation makes the image appear darker or closer to black and white.

[0096] (4) Blurring: Blurring is the process of applying a filter to an image to blur it, thereby simulating the effects of low light or unclear photography. Common blurring methods include Gaussian blur, which can be implemented using PIL.ImageFilter.GaussianBlur(). This operation helps simulate conditions such as low light or motion blur, making the model more robust to these conditions.

[0097] (5) Color Shifting: In low-light environments, the color of an image often shifts. This approach simulates color shifts in low-light environments by applying random shifts to the RGB channels of the image.

[0098] S3: Build a deep learning-based low-light substation road damage detection model. Input the paired substation road damage image dataset D3 into the low-light substation road damage detection model to identify the road damage type in the substation road damage image and automatically generate a positioning frame to determine the specific location and scope of the damage. Road damage types include cracks, potholes, warping, and settlement.

[0099] The specific steps of S3 are as follows:

[0100] S31: For any substation road damage image pair<p,p′> ∈D3, in the first branch, p is input into the first Conv module to generate the first normal light road damage feature map X1, and p′ is input into the first Conv module to generate the first low light road damage feature map X′1; in the second branch, p is input into the decomposition-based reflectivity and illumination decomposition module DRD (Decomposition-based Reflection and Illumination Decomposition) to generate the first normal light road damage reflectivity feature map R1, and p′ is input into the DRD module to generate the first low light road damage reflectivity feature map R′1; in order to simplify the description, X1 and X′1 constitute the first road damage feature map Figure 2 Tuple<X1,X′1> , R1 and R′1 constitute the first road damage reflectivity feature Figure 2 Tuple<R1,R′1> The first Conv module includes the following parts: a Conv2d layer with a convolution kernel size of 3×3 and a stride of 2, a BN layer, and a ReLU activation function, which are connected in series. The first Conv module can extract the edge and local features of the image, providing better input data for subsequent layers. The DRD module decomposes the reflectivity and illumination components of the image to extract key features while reducing the influence of ambient lighting.

[0101] Among them, the DRD module is an existing technology, which is mainly used to decompose images into reflectance and illumination components to generate high-quality pseudo labels. By effectively separating reflectance and illumination information, this module can improve the quality of training data, thereby improving the robustness and performance of the model. It is suitable for the scenarios applied by the present invention.

[0102] S32:<X1,X′1> Input into the second Conv module to generate the second road damage feature Figure 2Tuple<X2,X′2> , where X1 is input into the second Conv module to obtain the second normal light road damage feature map X2, and X′1 is input into the second Conv module to obtain the second low light road damage feature map X′2. Since the structure of the second Conv module is exactly the same as that of the first Conv module, the continuity and consistency of feature extraction can be ensured. On this basis,<X2,X′2> Passed to the CELAN module to generate a higher-level feature map, namely the third road damage feature Figure 2 Tuple<X3,X′3> , where X2 is input into the CELAN module to obtain the third normal-light road damage feature map X3, and X'2 is input into the CELAN module to obtain the third low-light road damage feature map X'3; the CELAN module is designed to enhance the global semantic information in the feature map and provide stronger expression capabilities for the low-light substation road damage detection model;

[0103] The present invention designs a cross-domain enhanced learning and adaptation network module CELAN (Cross-domain Enhanced Learning and Adaptation Network module); the CELAN module enhances the global feature capture capability of the low-light substation road damage detection model in low-light scenarios by combining multi-path, multi-scale feature extraction and frequency domain convolution, and improves the robustness of the low-light substation road damage detection model in low-light scenarios, thereby effectively improving the detection accuracy and generalization performance of the low-light substation road damage detection model; the overall structure of the CELAN module is as follows Figure 3 shown

[0104] The overall execution process of the CELAN module in S32 is as follows:

[0105] S321: In the first branch,<T1,T′1> Input to the Conv2d layer with a convolution kernel size of 3×3 for convolution operation to generate the second intermediate road damage feature Figure 2 Tuple<T2,T′2> , where T1 is input into the Conv2d layer with a convolution kernel size of 3×3 to obtain the second intermediate normal light road damage feature map T2, and T′1 is input into the Conv2d layer with a convolution kernel size of 3×3 to obtain the second intermediate low light road damage feature map T′2. This convolution operation can extract the basic texture information of the substation road damage image and improve the richness of feature expression; the step size of the Conv2d layer is 1, and the design of 0 padding is adopted to ensure<T2,T′2> The resolution remains unchanged, and then<T2,T′2> Input to the RepConv layer with a convolution kernel size of 3×3 for convolution operation to obtain the third intermediate road damage feature Figure 2Tuple<T3,T′3> , where T2 is input into the RepConv layer with a convolution kernel size of 3×3 to obtain the third intermediate normal light road damage feature map T3, and T′2 is input into the RepConv layer with a convolution kernel size of 3×3 to obtain the third intermediate low light road damage feature map T′3. The re-parameterized structure of the RepConv layer can effectively improve the speed of the convolution layer in the inference stage, while enhancing the nonlinear feature extraction capability in the training stage. Then<T3,T'3> Input to the FreqConv layer with a convolution kernel size of 5×5 for convolution operation to generate the fourth intermediate road damage feature Figure 2 Tuple<T4,T′4> , where T3 is input into the FreqConv layer with a convolution kernel size of 5×5 to obtain the fourth intermediate normal light road damage feature map T4, and T′3 is input into the FreqConv layer with a convolution kernel size of 5×5 to obtain the fourth intermediate low light road damage feature map T′4. In the FreqConv layer, the frequency domain feature extraction is used to further emphasize the structural characteristics of the damaged area, providing more recognizable information for substation road damage detection in low light scenarios.<T4,T′4> Input to the FreqConv layer with a convolution kernel size of 7×7 for convolution operation to generate the fifth intermediate road damage feature Figure 2 Tuple<T5,T′5> , where T4 is input into the FreqConv layer with a convolution kernel size of 7×7 to obtain the fifth intermediate normal light road damage feature map T5, and T′4 is input into the FreqConv layer with a convolution kernel size of 7×7 to obtain the fifth intermediate low light road damage feature map T′5. By further expanding the receptive field, the FreqConv layer can capture a wider range of context information, while taking into account the global feature expression and local feature expression of the damaged area. Finally,<T5,T′5> Input to the Conv2d layer with a convolution kernel size of 3×3 for convolution operation to generate the sixth intermediate road damage feature Figure 2 Tuple<T6,T′6> , where T5 is input into the Conv2d layer with a convolution kernel size of 3×3 to obtain the sixth intermediate normal light road damage feature map T6, and T′5 is input into the Conv2d layer with a convolution kernel size of 3×3 to obtain the sixth intermediate low light road damage feature map T′6;

[0106] After the above multiple convolution operations,<T6,T′6> The feature diversity and expression ability of the model have been improved, which strengthens the low-light substation road damage detection model's ability to characterize substation road damage and provides more substantial semantic information for subsequent feature extraction.

[0107] The reparameterized structure of the RepConv layer is a method for optimizing network inference efficiency. During the training phase, it uses a multi-branch structure (such as 3×3 convolution, 1×1 convolution, and skip connections) to enhance feature expression capabilities. During the inference phase, structural reparameterization is used to combine multiple branches into a standard 3×3 convolution, thereby reducing computational complexity and accelerating inference. This method improves inference efficiency while maintaining the model's expressiveness, making it more advantageous in computationally constrained environments.

[0108] S322: In the second branch,<T1,T′1> Input into a continuous convolution structure for dilated convolution operation to obtain the seventh intermediate road damage feature Figure 2 Tuple<T7,T′7> , wherein T1 is input into the continuous convolution structure to obtain the seventh intermediate normal light road damage feature map T7, and T′1 is input into the continuous convolution structure to obtain the seventh intermediate low light road damage feature map T′7; wherein the continuous convolution structure is composed of a Conv2d layer with a convolution kernel size of 1×1, a Conv2d layer with a convolution kernel size of 9×9, and a Conv2d layer with a convolution kernel size of 3×3, which are connected in series. The dilation rate of the Conv2d layer with a convolution kernel size of 1×1 and the Conv2d layer with a convolution kernel size of 9×9 are both 0, and the dilation rate of the Conv2d layer with a convolution kernel size of 3×3 is 9; the dilated convolution operation can effectively expand the receptive field, while capturing global features, maintaining the integrity of the detailed features of the substation road damage area;

[0109] S323:<T2,T′2> 、<T3,T'3> 、<T4,T′4> 、<T5,T′5> 、<T6,T′6> and

[0110] <T7,T′7> Perform Concat operation to generate the eighth intermediate road damage feature Figure 2 Tuple<T8,T′8> , where T2, T3, T4, T5, T6 and T7 are concat-operated to obtain the eighth intermediate normal light road damage feature map T8, and T′2, T′3, T′4, T′5, T′6 and T′7 are concat-operated to obtain the eighth intermediate low light road damage feature map T′8, thereby achieving the effective fusion of multi-scale features and different receptive field information, further enhancing the comprehensive expression ability of the low light substation road damage detection model for detail features and global features in low light scenes;

[0111] S324:<T8,T′8> The input is sent to the Conv2d layer with a convolution kernel size of 5×5 for convolution operation and activated by the Sigmoid activation function to generate the ninth middle road damage feature. Figure 2 Tuple<T9,T′9> , where T8 is input into the Conv2d layer with a convolution kernel size of 5×5 and activated by the Sigmoid activation function to obtain the ninth intermediate normal light road damage feature map T9, and T′8 is input into the Conv2d layer with a convolution kernel size of 5×5 and activated by the Sigmoid activation function to obtain the ninth intermediate low light road damage feature map T′9,<T9,T′9> As the final output of the CELAN module, it provides key input for subsequent feature extraction and prediction tasks;

[0112] Example:

[0113] Will<T1,T′1> Enter into two branches; in the first branch,<T1,T′1> Input to the Conv2d layer with a convolution kernel size of 3×3 for convolution operation to generate<T2,T′2> , where the sizes of T1 and T′1 are both 128×128×128, and the sizes of T2 and T′2 are both 128×128×256; then<T2,T′2> Input to the RepConv layer with a convolution kernel size of 3×3 for convolution operation, and get<T3,T'3> , where the sizes of T3 and T′3 are both 128×128×256; then<T3,T′3> Input to the FreqConv layer with a convolution kernel size of 5×5 for convolution operation to generate<T4,T′4> , where the sizes of T4 and T′4 are both 128×128×256;<T4,T′4> Input to the FreqConv layer with a convolution kernel size of 7×7 for convolution operation to generate<T5,T′5> , where the sizes of T5 and T′5 are both 128×128×256; finally<T5,T′5> Input to the Conv2d layer with a convolution kernel size of 3×3 for convolution operation to generate<T6,T′6> , where the sizes of T6 and T′6 are both 128×128×256;

[0114] In the second branch,<T1,T'1> Input into a continuous convolution structure for hole convolution operation, and get<T7,T'7> , where the sizes of T7 and T'7 are both 128×128×256;

[0115] Next,<T2,T'2> 、<T3,T'3> 、<T4,T'4> 、<T5,T'5> 、<T6,T'6> and<T7,T'7> Perform Concat operation to generate<T8,T'8> , where the sizes of T8 and T'8 are both 128×128×1536;

[0116] Finally,<T8,T'8> The input is sent to the Conv2d layer with a convolution kernel size of 5×5 for convolution operation and activated by the Sigmoid activation function to generate<T9,T'9> , where the sizes of T9 and T'9 are both 128×128×512;

[0117] S33:<X2,X′2> Input into the DLARK module to generate the second road damage reflectivity feature Figure 2 Tuple<R2,R′2> , where X2 is input into the DLARK module to obtain the second normal light road damage reflectance feature map R2, and X'1 is input into the DLARK module to obtain the second low light road damage reflectance feature map R'2; the DLARK module deeply extracts the reflectance features of the substation road damage image, capturing details and features, thereby providing input for the subsequent IRC module;

[0118] The present invention designs a dark light attenuation and reflection knowledge network module DLARK (Dark Light Attenuation and Reflection Knowledge Network module); the DLARK module is used for reflectance feature extraction, and by fusing multi-scale features and attention mechanism, the adaptability of the low-light substation road damage detection model in low-light conditions is improved; the DLARK module uses convolution operations with different void ratios to capture multi-scale information, enhance the perception of the target area, and at the same time combines the attention mechanism to focus on the key information area, effectively reduce the interference of background noise, and enhance the expression of detail features in complex scenes; by separately optimizing the representation of reflectance, the DLARK module can help the low-light substation road damage detection model extract inherent information related to the object itself from the image, avoid the interference of complex lighting conditions on target detection, thereby increasing the contribution of reflectance to the target detection task and reducing the negative impact of lighting information; at the same time, the module decouples the learning of reflectance features from the target detection task, better controls and strengthens the learning of reflectance information, and ultimately effectively improves the detection robustness and accuracy of the low-light substation road damage detection model in low-light scenes; the structure of the DLARK module is as follows Figure 4 As shown:

[0119] The overall execution process of the DLARK module in S33 is as follows:

[0120] S331: In the first branch,<X2,X'2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 9×9 and a void rate of 1 and a BN layer for convolution and normalization operations to obtain the first intermediate reflectivity feature. Figure 2 Tuple <X 2_1 ,X' 2_1>, where X2 is input into the convolution structure to obtain the first intermediate normal illumination reflectance feature map X 2_1 , input X'2 into the convolution structure to obtain the first intermediate low light reflectance feature map X' 2_1 ; This convolution operation captures a wide range of contextual information through a larger convolution kernel while ensuring the smoothness and consistency of feature expression;

[0121] S332: In the second branch,<X2,X'2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 5×5 and a void rate of 1 and a BN layer for convolution and normalization operations to generate the second intermediate reflectivity feature. Figure 2 Tuple <X 2_2 ,X′ 2_2 >, where X2 is input into the convolution structure to obtain the second intermediate normal illumination reflectance feature map X 2_2 , input X′2 into the convolution structure to obtain the second intermediate low light reflectance feature map X′ 2_2 The smaller convolution kernel in this convolution operation focuses on extracting detailed features, providing important support for subsequent multi-scale fusion.

[0122] S333: In the third branch,<X2,X'2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 5×5 and a void rate of 2 and a BN layer for convolution and normalization operations to obtain the third intermediate reflectivity feature. Figure 2 Tuple <X 2_3 ,X' 2_3 >, where X2 is input into the convolution structure to obtain the third intermediate normal illumination reflectance feature map X 2_3 , input X'2 into the convolution structure to obtain the third intermediate low light reflectance feature map X' 2_3 The increase in the dilation rate in the convolution operation will further expand the receptive field and effectively capture the correlation information between long-distance features.

[0123] S334: In the fourth branch,<X2,X'2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 2×2 and a void rate of 2 and a BN layer for convolution and normalization operations to generate the fourth intermediate reflectivity feature. Figure 2 Tuple <X 2_4 ,X' 2_4 >, where X2 is input into the convolution structure to obtain the fourth intermediate normal illumination reflectance feature map X 2_4 , input X′2 into the convolution structure to obtain the fourth intermediate low light reflectance feature map X' 2_4 ;

[0124] S335: In the fifth branch,<X2,X'2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 2×2 and a void rate of 4 and a BN layer for convolution and normalization operations to obtain the fifth intermediate reflectivity feature. Figure 2 Tuple <X 2_5 ,X' 2_5 >, where X2 is input into the convolution structure to obtain the fifth intermediate normal illumination reflectance feature map X 2_5 , input X'2 into the convolution structure to obtain the fifth intermediate low light reflectance feature map X' 2_5 Through the design of multiple void ratios, the DLARK module has the ability to capture features at different scales, which is especially effective in low-light scenes.

[0125] S336: Yes <X 2_1 ,X' 2_1 >、 <X 2_2 ,X′ 2_2 >、 <X 2_3 ,X' 2_3 >、 <X 2_4 ,X' 2_4 > and

[0126] <X 2_5 ,X' 2_5 >Perform the Concat operation to obtain the sixth intermediate reflectivity feature Figure 2 Tuple <X 2_6 ,X' 2_6 >, where X 2_1 、X 2_2 、X 2_3 、X 2_4 and X 2_5 Perform the Concat operation to obtain the sixth intermediate normal light reflectance feature map X 2_6 , X' 2_1 、X' 2_2 、X' 2_3 、X' 2_4 and X' 2_5 Perform the Concat operation to obtain the sixth intermediate low light reflectance feature map X' 2_6 , the Concat operation integrates multi-scale information to make the feature expression more comprehensive; <X 2_6 ,X' 2_6 > Input to the AvgPool layer for pooling operation to obtain the seventh intermediate reflectivity feature Figure 2 Tuple <X 2_7 ,X' 2_7 >, where X 2_6 Input to the AvgPool layer to obtain the seventh intermediate normal light reflectance feature map X 2_7 , X'2_6 Input to the AvgPool layer to obtain the seventh intermediate low light reflectance feature map X' 2_7 , the pooling operation enhances the globality of features by downsampling, while reducing the impact of redundant features. <X 2_7 ,X' 2_7 > Input to the continuous convolution structure consisting of a Conv2d layer with a convolution kernel size of 1×1, a DWConv layer with a convolution kernel size of 2×2, a DWConv layer with a convolution kernel size of 5×5, and a Conv2d layer with a convolution kernel size of 2×2 for deep convolution operation to generate the eighth intermediate reflectivity feature Figure 2 Tuple <X 2_8 ,X' 2_8 >, where X 2_7 Input to the continuous convolution structure to obtain the eighth intermediate normal light reflectance feature map X 2_8 , X′ 2_7 Input to the continuous convolution structure to obtain the eighth intermediate low light reflectance feature map X' 2_8 The introduction of deep convolution greatly reduces the amount of computation while retaining key features. <X 2_6 ,X′ 2_6 >While performing the pooling operation, <X 2_6 ,X' 2_6 > Input to the BN layer for normalization operation to generate the ninth intermediate reflectivity feature Figure 2 Tuple <X 2_9 ,X' 2_9 >, where X 2_6 Input to the BN layer to obtain the ninth intermediate normal light reflectance feature map X 2_9 , X' 2_6 Input to the BN layer to obtain the ninth intermediate low light reflectance feature map X' 2_9 ,The normalization operation avoids the problem of gradient disappearance or explosion, thereby improving the training efficiency, and plays a regularization role to prevent overfitting;

[0127] S337: <X 2_9 ,X' 2_9 > Input to the AvgPool layer for pooling operation to obtain the tenth intermediate reflectivity feature Figure 2 Tuple <X 2_10 ,X' 2_10 >, where X 2_9 Input to the AvgPool layer to obtain the tenth intermediate normal light reflectance feature map X 2_10 , X' 2_9 Input to the AvgPool layer to obtain the tenth intermediate low light reflectance feature map X' 2_10 , then <X 2_10 ,X'2_10 > Input into the structure consisting of a fully connected layer, a ReLU activation function and a fully connected layer sequence to perform feature transformation operations to obtain the eleventh intermediate reflectivity feature Figure 2 Tuple <X 2_11 ,X′ 2_11 >, where X 2_10 Input to the structure composed of the full connection layer, ReLU activation function and full connection layer sequence to obtain the eleventh intermediate normal light reflectance feature map X 2_11 , X' 2_10 Input to the structure composed of a fully connected layer, a ReLU activation function and a fully connected layer in sequence to obtain the eleventh intermediate low light reflectance feature map X' 2_11 , the feature transformation operation further integrates the pooled features, improves the expression ability of the module, and provides a more abstract feature representation for subsequent activation and decision-making processes. <X 2_11 ,X' 2_11 >Activate through the Sigmoid activation function to obtain the twelfth intermediate reflectivity feature Figure 2 Tuple <X 2_12 ,X' 2_12 >, where X 2_11 The twelfth intermediate normal illumination reflectance feature map X is obtained by activating the Sigmoid function 2_12 , X' 2_11 The twelfth intermediate low-light reflectance feature map X′ is obtained by activating the Sigmoid function 2_12 ,Will <X 2_9 ,X′ 2_9 >with <X 2_12 ,X' 2_12 >Perform element-by-element multiplication to obtain the thirteenth intermediate reflectivity feature Figure 2 Tuple <X 2_13 ,X' 2_13 >, where X 2_9 With X 2_12 Perform element-by-element multiplication to obtain the thirteenth intermediate normal light reflectance feature map X 2_13 , X' 2_9 With X' 2_12 Perform element-by-element multiplication to obtain the thirteenth intermediate low-light reflectance feature map X' 2_13 , the element-by-element multiplication operation introduces interactive features, <X 2_9 ,X′ 2_9 > and <X 2_12 ,X' 2_12 > information to enhance the feature expression ability. Element-by-element multiplication can strengthen specific features and suppress irrelevant parts, thereby enhancing the module's sensitivity to important features and improving the accuracy of the final output. <X 2_13,X' 2_13 > Input into the structure consisting of a fully connected layer, a ReLU activation function and a fully connected layer sequence to perform feature transformation operation, and obtain the fourteenth intermediate reflectivity feature Figure 2 Tuple <X 2_14 ,X' 2_14 >, where X 2_13 Input to the structure composed of the full connection layer, ReLU activation function and full connection layer sequence to obtain the fourteenth intermediate normal light reflectance feature map X 2_14 , X' 2_13 Input to the structure composed of a fully connected layer, a ReLU activation function and a fully connected layer in sequence to obtain the fourteenth intermediate low light reflectance feature map X′ 2_14 ;

[0128] S338: Yes <X 2_14 ,X' 2_14 >After performing the DropPath operation, the fifteenth intermediate reflectivity feature is obtained Figure 2 Tuple <X 2_15 ,X' 2_15 >, where X 2_14 Perform DropPath operation to obtain the fifteenth intermediate normal light reflectance feature map X 2_15 , X' 2_14 Perform DropPath operation to obtain the fifteenth intermediate low light reflectance feature map X' 2_15 , the DropPath operation increases the robustness of the module and reduces overfitting by randomly dropping the path. <X 2_15 ,X' 2_15 >with <X 2_8 ,X' 2_8 >Perform the Concat operation to obtain the final second road damage reflectivity feature Figure 2 Tuple<R2,R'2> , where X 2_15 With X 2_8 Perform Concat operation to obtain the second normal light road damage reflectivity feature map R2, and convert X' 2_15 With X' 2_8 The Concat operation is performed to obtain the second low-light road damage reflectivity feature map R'2. Figure 2 The tuple, as the final output of the DLARK module, provides key input for subsequent feature extraction and prediction tasks;

[0129] DropPath is a regularization method that randomly discards some paths during neural network training, forcing the network to learn more robust feature representations. In this way, DropPath avoids the model's over-reliance on certain specific paths, enhances the network's generalization ability, and improves its performance on unseen data. It is particularly suitable for deep networks and can effectively improve model stability and performance during training.

[0130] Example:

[0131] The second road damage feature Figure 2 Tuple<X2,X'2> Input to five branches, where the size of X2 and X'2 are both 128×128×128;

[0132] In the first branch,<X2,X'2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 9×9 and a void rate of 1 and a BN layer for convolution and normalization operations to obtain <X 2_1 ,X' 2_1 >, where X 2_1 and X′ 2_1 The size of is 128×128×256; in the second branch,<X2,X'2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 5×5 and a void rate of 1 and a BN layer for convolution and normalization operations to generate <X 2_2 ,X' 2_2 >, where X 2_2 and X' 2_2 The size of is 128×128×128; in the third branch,<X2,X'2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 5×5 and a void rate of 2 and a BN layer for convolution and normalization operations to obtain <X 2_3 ,X' 2_3 >, where X 2_3 and X' 2_3 The size of is 128×128×128; in the fourth branch,<X2,X'2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 2×2 and a void rate of 2 and a BN layer for convolution and normalization operations to generate <X 2_4 ,X' 2_4 >, where X 2_4 and X' 2_4 The size of is 128×128×128; in the fifth branch,<X2,X'2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 2×2 and a void rate of 4 and a BN layer for convolution and normalization operations to obtain <X 2_5 ,X' 2_5>, where X 2_5 and X' 2_5 The size is 128×128×128;

[0133] Next, <X 2_1 ,X' 2_1 >、 <X 2_2 ,X' 2_2 >、 <X 2_3 ,X' 2_3 >、 <X 2_4 ,X' 2_4 > and <X 2_5 ,X' 2_5 >Perform Concat operation to get <X 2_6 ,X' 2_6 >, where X 2_6 and X' 2_6 The size of is 128×128×768; then, <X 2_6 ,X' 2_6 > Input to the AvgPool layer for pooling operation, and get <X 2_7 ,X' 2_7 >, where X 2_7 and X' 2_7 The size of is 128×128×768; then, <X 2_7 ,X' 2_7 > Input to the continuous convolution structure consisting of a Conv2d layer with a convolution kernel size of 1×1, a DWConv layer with a convolution kernel size of 3×3, a DWConv layer with a convolution kernel size of 5×5, and a Conv2d layer with a convolution kernel size of 2×2 for depth convolution operation to generate <X 2_8 ,X' 2_8 >, where X 2_8 and X' 2_8 The size of the <X 2_6 ,X' 2_6 >While performing the pooling operation, <X 2_6 ,X' 2_6 > Input to the BN layer for normalization operation to generate <X 2_9 ,X' 2_9 >, where X 2_9 and X' 2_9 The size is 64×64×256;

[0134] Will <X 2_9 ,X' 2_9 > Input to the AvgPool layer for pooling operation, and get <X 2_10 ,X' 2_10 >, where X 2_10 and X'2_10 The size of is 64×64×128; then <X 2_10 ,X' 2_10 > Input into the structure consisting of a fully connected layer, a ReLU activation function and a fully connected layer sequence for feature transformation operation, and obtain <X 2_11 ,X' 2_11 >, where X 2_11 and X' 2_11 The size of X is 64×64×512; 2_11 Activate through Sigmoid activation function and get <X 2_12 ,X' 2_12 >, where X 2_12 and X' 2_12 The size of is 64×64×512; next, <X 2_9 ,X' 2_9 >with <X 2_12 ,X' 2_12 >Perform element-wise multiplication to get <X 2_13 ,X' 2_13 >, where X 2_13 and X' 2_13 The size of is 64×64×256; then, <X 2_13 ,X' 2_13 > Input into the structure consisting of a fully connected layer, a ReLU activation function and a fully connected layer sequence for feature transformation operation, and obtain <X 2_14 ,X' 2_14 >, where X 2_14 and X' 2_14 The size of the 64 × 64 × 128; <X 2_14 ,X' 2_14 >After performing the DropPath operation, we get <X 2_15 ,X' 2_15 >, where X 2_15 and X' 2_15 The size of is 64×64×128; finally, <X 2_15 ,X' 2_15 >with <X 2_8 ,X' 2_8 >Perform Concat operation to get<R2,R'2> , where the sizes of R2 and R'2 are both 64×64×640;

[0135] S34: In order to further distinguish the reflectance and illumination components in the image and maintain the consistency between illumination and reflectance,<R1,R'1> and<R2,R′2> Input into the Illumination and Reflection Compensation module IRC (Illumination and Reflection Compensation), and generate the third road damage reflectivity feature by introducing the process of interaction, reorganization and consistency constraints. Figure 2 Tuple<R3,R'3> , wherein R1 and R2 are input into the IRC module to obtain the third normal light road damage reflectivity characteristic map R3, and R'1 and R'2 are input into the IRC module to obtain the third low light road damage reflectivity characteristic map R'3;

[0136] S35: Yes<R3,R′3> and<X3,X′3> Perform the Concat operation to fuse the reflectivity feature with the high-level semantic feature to obtain the fourth road damage feature. Figure 2 Tuple<X4,X'4> , wherein R3 and X3 are concat-operated to obtain the fourth normal-light road damage feature map X4, and R′3 and X′3 are concat-operated to obtain the fourth low-light road damage feature map X′4;

[0137] S36:<X4,X'4> This data is input into the target prediction head, generating a tensor containing prediction information. Each row of the tensor corresponds to a detection result, including the predicted bounding box coordinates, damage category label, and confidence score information. The target prediction head achieves accurate detection and classification of substation road damage images through layered extraction, feature fusion, and consistency modeling. It also improves the adaptability and robustness of the low-light substation road damage detection model in low-light environments.

[0138] In deep learning object detection models, the Head module is responsible for in-depth analysis and processing of feature maps to predict the object's category, location, and confidence level. The output of the Head module is typically a three-dimensional tensor, with each row containing prediction information at a specific spatial location, including the category probability, bounding box coordinates, and confidence level of the object for each predicted box. This information enables the model to accurately locate and classify objects.

[0139] (1) Bounding box prediction

[0140] In object detection tasks, images are usually divided into multiple grids, and each grid cell is responsible for predicting several bounding boxes. The bounding box prediction task is to determine the location of the object in the image, which mainly includes the following parameters:

[0141] Center point coordinates (x, y): relative to the center point of the feature map cell, the position of the bounding box center in the entire image is determined by offset calculation;

[0142] Width and height (w,h): The width and height of the bounding box, expressed relative to the width and height of the entire image to ensure that the scale and position of the predicted box are accurate;

[0143] To stabilize training, bounding box coordinates are usually normalized using a Sigmoid function to limit the predicted values to between 0 and 1. This normalization helps improve the accuracy of coordinate predictions and avoids box sizes that are too large or too small, thereby reducing prediction bias.

[0144] (2) Category and confidence prediction

[0145] In the prediction of each bounding box, in addition to the coordinates, the category and confidence of the object need to be predicted:

[0146] Category prediction: The model outputs a probability vector representing the probability of each category existing within the bounding box. Using this probability, the model can identify the specific type of object, such as an insulator or a pole.

[0147] Confidence prediction: The confidence level indicates the model's confidence in the existence of the object in the current bounding box. Its value ranges from 0 to 1, with larger values indicating greater confidence in the bounding box containing the target object. In post-processing, high-confidence bounding boxes are usually filtered out based on a set confidence threshold to improve the accuracy of the detection results.

[0148] In practice, the Head module outputs a three-dimensional tensor containing detailed information about all predicted boxes at each spatial location, namely, class probability, bounding box coordinates, and confidence. For each location in the feature map, the tensor provides the prediction information for the corresponding bounding box. This design enables the model to output both classification and localization information, thereby improving detection accuracy and efficiency.

[0149] Example:

[0150] Assume that the size of the feature map is H×W×C, where H is the height of the feature map, W is the width of the feature map, and C is the number of channels of the feature map;

[0151] The tuple<p,p'> ∈D3 is input to the first Conv module to generate<X1,X'1> , where the sizes of p and p' are both 512×512×3, and the sizes of X1 and X'1 are both 256×256×64;<p,p'> ∈D3 is input to the DRD module to generate<R1,R'1> , where the sizes of R1 and R'1 are both 64×64×640;

[0152] Next,<X1,X'1> Input into the second Conv module to generate<X2,X'2> , where the sizes of X2 and X'2 are both 128×128×128;<X2,X'2> Passed to the CELAN module to generate<X3,X'3> , where the sizes of X3 and X'3 are both 128×128×512;<X2,X'2> When passing it to the CELAN module,<X2,X'2> Input into the DLARK module to generate<R2,R'2> , where the sizes of R2 and R'2 are both 64×64×640;

[0153] In order to further distinguish the reflectance and illumination components in the image and maintain the consistency between illumination and reflectance,<R1,R'1> and<R2,R'2> Input to the IRC module to generate<R3,R'3> , where the sizes of R3 and R'3 are both 128×128×256;

[0154] Then,<R3,R'3> and<X3,X'3> Perform Concat operation to fuse reflectivity features with high-level semantic features to obtain<X4,X'4> , where the sizes of X4 and X'4 are both 128×128×768;

[0155] Finally,<X4,X'4> This is input into the target prediction head, generating a tensor containing the prediction information. Each row of the tensor corresponds to a detection result, including the predicted bounding box coordinates, damage category label, and confidence score. Through layered extraction, feature fusion, and consistency modeling, the entire process achieves accurate detection and classification of substation road damage images, while also improving the model's adaptability and robustness in low-light environments.

[0156] S4: Before training the low-light substation road damage detection model, you first need to set hyperparameters, such as selecting an optimizer, determining the initial learning rate, batch size, and number of training rounds. After completing the initial configuration, the paired substation road damage image dataset is divided into a training set and a validation set to ensure the independence and representativeness of each dataset.

[0157] The specific steps of S4 are as follows:

[0158] S41: During the training process, the model extracts data from the training set according to the set batch size for forward calculation to generate prediction results. Subsequently, the model performs backpropagation through the overall loss function L2R-LOSS to optimize the parameters of the substation road damage detection model in low-light environments.

[0159] This paper designs a L2R-LOSS loss function for training a low-light substation road damage detection model, addressing the issue of decreased target detection accuracy under low-light conditions. The L2R-LOSS loss function is used to enhance the model's target feature extraction capability in low-light environments, thereby improving detection stability and accuracy. By integrating reflectivity learning loss, reorganization consistency loss, and target detection loss, the collaborative optimization of multiple losses is achieved, reducing the impact of illumination changes on detection results, enhancing the model's adaptability in low-light conditions, and improving the robustness of substation road damage detection.

[0160] The loss function L2R-LOSS in S41 is composed of the reflectivity learning loss function L r , Reorganization consistency loss function L rc And the target detection loss function L target The weighted composition, its function expression is as follows:

[0161] L L2R-LOSS =λ g L r +λ f L target +λ q L rc

[0162] Among them, λ g ,λ f and λ q is the loss weight, the sum of the three is 1;

[0163] Among them, L r The function expression is as follows:

[0164]

[0165] Among them, R'1 is the first low-light road damage reflectivity feature map; R'2 is the first low-light road damage reflectivity feature map; N is the total number of feature points of the R'1 and R'2 feature maps, R' 1_i and R' 2_i Respectively represent the feature response value corresponding to the i-th position in the feature map; and denote the average characteristic response values of R'1 and R'2, respectively, and are the variances of the characteristic responses R'1 and R'2, is the covariance of the characteristic response values between R'1 and R'2, and C1 and C2 are constants to avoid the denominator being zero;

[0166] Among them, L target The function expression is as follows:

[0167]

[0168] Among them, α and β represent the weights of position regression loss and classification loss respectively; x i and x' i Represents the center point of the real box and the predicted box of the i-th target respectively; p i and p' i They represent the true category label and predicted category probability of the i-th target, respectively, and Smooth(x) represents the error between the predicted value and the true value, that is,

[0169]

[0170] L rc The function expression is as follows:

[0171] L rc =||R1-R′2||1+||R′1-R2||1

[0172] Among them, R1 is the first normal light road damage reflectivity characteristic map; R'2 is the second low light road damage reflectivity characteristic map; R'1 is the first low light road damage reflectivity characteristic map; R2 is the second normal light road damage reflectivity characteristic map;

[0173] S42: After each round of training, the validation set data is used to evaluate the model to determine the generalization ability of the model;

[0174] S43: To prevent overfitting of the model during training, an early stopping mechanism is introduced during training. If the loss on the validation set does not improve significantly over multiple consecutive training rounds, training will be terminated early. This helps save computing resources and ensures the stability of the model in practical applications. Overfitting refers to the situation where the model performs well on the training set but performs poorly on the validation set.

[0175] S44: When the training reaches the predetermined number of rounds and the validation set loss is stable or reaches an ideal level, the training process ends, and an optimized low-light substation road damage detection model is finally obtained. The model can accurately detect substation road damage under low-light conditions and maintain good performance in different environments;

[0176] S45: After training, the model will have appropriate parameter configuration and strong generalization ability, and can be applied to the new substation road damage detection task;

[0177] S5: After model training is completed, the trained low-light substation road damage detection model is used to identify and analyze the current substation road damage data. The final output includes the specific substation road damage type and location identification results, providing substation inspection personnel with accurate early warning information and subsequent processing suggestions.

[0178] Although the present invention has been described above with reference to embodiments, various modifications may be made thereto and equivalent components may be substituted without departing from the scope of the present invention. In particular, as long as there are no structural conflicts, the various features of the embodiments disclosed herein may be combined with each other in any manner, and the omission of an exhaustive description of such combinations in this specification is solely for the sake of space and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A low-light substation road damage detection method based on deep learning, characterized in that: The specific steps are as follows: S1: Construct the original substation road damage image dataset D1; S2: Preprocess D1 to obtain the processed substation road damage image dataset D2, and perform low-light degradation synthesis operation on each substation road damage image p in D2 to obtain a low-light version image p', and form a substation road damage image binary<p,p'> , a paired substation road damaged image dataset D3 is constructed from all substation road damaged image pairs; S3: Build a deep learning-based low-light substation road damage detection model. Input the paired substation road damage image dataset D3 into the low-light substation road damage detection model to identify the road damage type in the substation road damage image and automatically generate a positioning frame to determine the specific location and scope of the damage. Road damage types include cracks, potholes, warping, and settlement. S4: Before training the low-light substation road damage detection model, you first need to set hyperparameters. After completing the initial configuration, the paired substation road damage image dataset is divided into a training set and a validation set to ensure the independence and representativeness of each dataset. S5: After model training is completed, the trained low-light substation road damage detection model is used to identify and analyze the current substation road damage data. The final output includes the specific substation road damage type and location identification results, providing substation inspection personnel with accurate early warning information and subsequent processing suggestions.

2. The method for detecting road damage in low-light substations based on deep learning according to claim 1, characterized in that: The specific steps of S1 are as follows: S11: Collect images of damaged substation roads under normal lighting to form an original substation road damaged image dataset D1; S12: Manually label D1 to provide labels for subsequent model training. During the labeling process, use the labeling tool to select the road damage area and add corresponding labels to each damaged area to mark the road damage type in the area.

3. The method for detecting road damage in low-light substations based on deep learning according to claim 1, characterized in that: The specific steps of S2 are as follows: S21: Crop all substation road damage images in D1 to a uniform size; S22: Expand D1 by using data augmentation operations such as rotation, flipping, and noise addition to generate a substation road damage image dataset D2 containing more substation road damage images; S23: Process each substation road damage image p in D2 by using brightness adjustment, contrast adjustment, color saturation reduction, blurring, and color shifting operations to generate a corresponding low-light version image p'; S24: Combine p and p' to form a substation road damage image binary<p,p'> , a paired substation road damaged image dataset D3 is constructed by all the above substation road damaged image binary pairs.

4. The method for detecting road damage in low-light substations based on deep learning according to claim 1, characterized in that: The specific steps of S3 are as follows: S31: For any substation road damage image pair<p,p'> ∈D3, in the first branch, p is input into the first Conv module to generate the first normal light road damage feature map X1, and p' is input into the first Conv module to generate the first low light road damage feature map X'1; in the second branch, p is input into the decomposition-based reflectivity and illumination decomposition module DRD to generate the first normal light road damage reflectivity feature map R1, and p' is input into the DRD module to generate the first low light road damage reflectivity feature map R'1; S32:<X1,X'1> Input into the second Conv module to generate the second road damage feature map binary<X2,X'2> , where X1 is input into the second Conv module to obtain the second normal light road damage feature map X2, and X'1 is input into the second Conv module to obtain the second low light road damage feature map X'2. Since the structure of the second Conv module is exactly the same as that of the first Conv module, the continuity and consistency of feature extraction can be ensured. On this basis,<X2,X'2> Passed to the CELAN module to generate a higher-level feature map, namely the third road damage feature map binary<X3,X'3> , wherein X2 is input into the CELAN module to obtain the third normal-light road damage feature map X3, and X'2 is input into the CELAN module to obtain the third low-light road damage feature map X'3; S33:<X2,X'2> Input into the DLARK module to generate the second road damage reflectivity feature map binary<R2,R'2> , wherein X2 is input into the DLARK module to obtain the second normal light road damage reflectivity feature map R2, and X'1 is input into the DLARK module to obtain the second low light road damage reflectivity feature map R'2; S34: In order to further distinguish the reflectance and illumination components in the image and maintain the consistency between illumination and reflectance,<R1,R'1> and<R2,R'2> Input into the illumination and reflection compensation module IRC, and generate the third road damage reflectivity feature map binary by introducing the process of interaction, recombination and consistency constraints<R3,R'3> , wherein R1 and R2 are input into the IRC module to obtain a third normal light road damage reflectivity characteristic map R3, and R'1 and R'2 are input into the IRC module to obtain a third low light road damage reflectivity characteristic map R'3; S35: Yes<R3,R'3> and<X3,X'3> Perform the Concat operation to fuse the reflectivity feature and the high-level semantic feature to obtain the fourth road damage feature map binary.<X4,X'4> , wherein R3 and X3 are concat-operated to obtain the fourth normal-light road damage feature map X4, and R'3 and X'3 are concat-operated to obtain the fourth low-light road damage feature map X'4; S36:<X4,X'4> The input is fed into the target prediction head to generate a tensor containing the prediction information. Each row of the tensor corresponds to a detection result, including the predicted bounding box coordinates, damage category label, and confidence score information.

5. The method for detecting road damage in low-light substations based on deep learning according to claim 4, characterized in that: The overall execution process of the CELAN module in S32 is as follows: S321: In the first branch,<T1,T'1> The input is sent to the Conv2d layer with a convolution kernel size of 3×3 for convolution operation to generate the second intermediate road damage feature map binary.<T2,T'2> , where T1 is input into the Conv2d layer with a convolution kernel size of 3×3 to obtain the second intermediate normal light road damage feature map T2, and T'1 is input into the Conv2d layer with a convolution kernel size of 3×3 to obtain the second intermediate low light road damage feature map T'2, and then<T2,T'2> The input is sent to the RepConv layer with a convolution kernel size of 3×3 for convolution operation to obtain the third middle road damage feature map binary.<T3,T'3> , where T2 is input to the RepConv layer with a convolution kernel size of 3×3 to obtain the third intermediate normal light road damage feature map T3, and T'2 is input to the RepConv layer with a convolution kernel size of 3×3 to obtain the third intermediate low light road damage feature map T'3, and then<T3,T'3> Input to the FreqConv layer with a convolution kernel size of 5×5 for convolution operation to generate the fourth middle road damage feature map binary<T4,T'4> , where T3 is input into the FreqConv layer with a convolution kernel size of 5×5 to obtain the fourth middle normal light road damage feature map T4, and T'3 is input into the FreqConv layer with a convolution kernel size of 5×5 to obtain the fourth middle low light road damage feature map T'4, and then<T4,T'4> The convolution operation is performed on the FreqConv layer with a convolution kernel size of 7×7 to generate the fifth middle road damage feature map binary.<T5,T'5> , where T4 is input into the FreqConv layer with a convolution kernel size of 7×7 to obtain the fifth middle normal light road damage feature map T5, and T'4 is input into the FreqConv layer with a convolution kernel size of 7×7 to obtain the fifth middle low light road damage feature map T'5.<T5,T'5> Input to the Conv2d layer with a convolution kernel size of 3×3 for convolution operation to generate the sixth intermediate road damage feature map binary<T6,T'6> , wherein T5 is input into the Conv2d layer with a convolution kernel size of 3×3 to obtain the sixth intermediate normal light road damage feature map T6, and T'5 is input into the Conv2d layer with a convolution kernel size of 3×3 to obtain the sixth intermediate low light road damage feature map T'6; S322: In the second branch,<T1,T'1> Input into a continuous convolution structure for dilated convolution operation to obtain the seventh intermediate road damage feature map binary<T7,T'7> , wherein T1 is input into the continuous convolution structure to obtain the seventh intermediate normal light road damage feature map T7, and T'1 is input into the continuous convolution structure to obtain the seventh intermediate low light road damage feature map T'7; S323:<T2,T'2> 、<T3,T'3> 、<T4,T'4> 、<T5,T'5> 、<T6,T'6> and <T7,T'7> Perform the Concat operation to generate the eighth middle road damage feature map binary<T8,T'8> , where T2, T3, T4, T5, T6 and T7 are concat-operated to obtain the eighth intermediate normal light road damage feature map T8, and T'2, T'3, T'4, T'5, T'6 and T'7 are concat-operated to obtain the eighth intermediate low light road damage feature map T'8, thereby achieving the effective fusion of multi-scale features and different receptive field information, further enhancing the comprehensive expression ability of the low light substation road damage detection model for detail features and global features in low light scenes; S324:<T8,T'8> The input is sent to the Conv2d layer with a convolution kernel size of 5×5 for convolution operation and activated by the Sigmoid activation function to generate the ninth middle road damage feature map binary.<T9,T'9> , where T8 is input into the Conv2d layer with a convolution kernel size of 5×5 and activated by the Sigmoid activation function to obtain the ninth middle normal light road damage feature map T9, and T'8 is input into the Conv2d layer with a convolution kernel size of 5×5 and activated by the Sigmoid activation function to obtain the ninth middle low light road damage feature map T'9,<T9,T'9> As the final output of the CELAN module, it provides key input for subsequent feature extraction and prediction tasks.

6. The method for detecting road damage in low-light substations based on deep learning according to claim 4, characterized in that: The overall execution process of the DLARK module in S33 is as follows: S331: In the first branch,<X2,X'2> Input to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 9×9 and a void rate of 1 and a BN layer for convolution and normalization operations to obtain the first intermediate reflectivity feature map binary <X 2_1 ,X' 2_1 >, where X2 is input into the convolution structure to obtain the first intermediate normal illumination reflectance feature map X 2_1 , input X'2 into the convolution structure to obtain the first intermediate low light reflectance feature map X' 2_1 ; S332: In the second branch,<X2,X'2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 5×5 and a void rate of 1 and a BN layer for convolution and normalization operations to generate the second intermediate reflectivity feature map binary. <X 2_2 ,X' 2_2 >, where X2 is input into the convolution structure to obtain the second intermediate normal illumination reflectance feature map X 2_2 , input X'2 into the convolution structure to obtain the second intermediate low light reflectance feature map X' 2_2 ; S333: In the third branch,<X2,X'2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 5×5 and a void rate of 2 and a BN layer for convolution and normalization operations to obtain the third intermediate reflectivity feature map binary. <X 2_3 ,X' 2_3 >, where X2 is input into the convolution structure to obtain the third intermediate normal illumination reflectance feature map X 2_3 , input X'2 into the convolution structure to obtain the third intermediate low light reflectance feature map X' 2_3 ; S334: In the fourth branch,<X2,X′2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 2×2 and a void rate of 2 and a BN layer for convolution and normalization operations to generate the fourth intermediate reflectivity feature map binary. <X 2_4 ,X' 2_4 >, where X2 is input into the convolution structure to obtain the fourth intermediate normal illumination reflectance feature map X 2_4 , input X'2 into the convolution structure to obtain the fourth intermediate low light reflectance feature map X' 2_4 ; S335: In the fifth branch,<X2,X'2> The input is sent to the convolution structure consisting of a Conv2d layer with a convolution kernel size of 2×2 and a void rate of 4 and a BN layer for convolution and normalization operations to obtain the fifth intermediate reflectivity feature map binary. <X 2_5 ,X' 2_5 >, where X2 is input into the convolution structure to obtain the fifth intermediate normal illumination reflectance feature map X 2_5 , input X'2 into the convolution structure to obtain the fifth intermediate low light reflectance feature map X' 2_5 ; S336: Yes <X 2_1 ,X' 2_1 >、 <X 2_2 ,X' 2_2 >、 <X 2_3 ,X' 2_3 >、 <X 2_4 ,X' 2_4 > and <X 2_5 ,X' 2_5 >Perform the Concat operation to obtain the sixth intermediate reflectivity feature map binary <X 2_6 ,X' 2_6 >, where X 2_1 、X 2_2 、X 2_3 、X 2_4 and X 2_5 Perform the Concat operation to obtain the sixth intermediate normal light reflectance feature map X 2_6 , X' 2_1 、X' 2_2 、X' 2_3 、X' 2_4 and X' 2_5 Perform the Concat operation to obtain the sixth intermediate low light reflectance feature map X' 2_6 ,Will <X 2_6 ,X' 2_6 > Input to the AvgPool layer for pooling operation to obtain the seventh intermediate reflectivity feature map binary <X 2_7 ,X' 2_7 >, where X 2_6 Input to the AvgPool layer to obtain the seventh intermediate normal light reflectance feature map X 2_7 , X' 2_6 Input to the AvgPool layer to obtain the seventh intermediate low light reflectance feature map X' 2_7 ,Will <X 2_7 ,X' 2_7 > Input to the continuous convolution structure consisting of a Conv2d layer with a convolution kernel size of 1×1, a DWConv layer with a convolution kernel size of 2×2, a DWConv layer with a convolution kernel size of 5×5, and a Conv2d layer with a convolution kernel size of 2×2 for deep convolution operation to generate the eighth intermediate reflectivity feature map binary <X 2_8 ,X' 2_8 >, where X 2_7 Input to the continuous convolution structure to obtain the eighth intermediate normal light reflectance feature map X 2_8 , X′ 2_7 Input to the continuous convolution structure to obtain the eighth intermediate low light reflectance feature map X′ 2_8 , in the right <X 2_6 ,X' 2_6 >While performing the pooling operation, <X 2_6 ,X' 2_6 > Input to the BN layer for normalization operation to generate the ninth intermediate reflectivity feature map binary <X 2_9 ,X' 2_9 >, where X 2_6 Input to the BN layer to obtain the ninth intermediate normal light reflectance feature map X 2_9 , X' 2_6 Input to the BN layer to obtain the ninth intermediate low light reflectance feature map X' 2_9 ; S337: <X 2_9 ,X' 2_9 > Input to the AvgPool layer for pooling operation to obtain the tenth intermediate reflectivity feature map binary <X 2_10 ,X' 2_10 >, where X 2_9 Input to the AvgPool layer to obtain the tenth intermediate normal light reflectance feature map X 2_10 , X' 2_9 Input to the AvgPool layer to obtain the tenth intermediate low light reflectance feature map X' 2_10 , then <X 2_10 ,X' 2_10 > Input into the structure consisting of a fully connected layer, a ReLU activation function and a fully connected layer sequence to perform feature transformation operations, and obtain the eleventh intermediate reflectivity feature map binary <X 2_11 ,X' 2_11 >, where X 2_10 Input to the structure composed of the full connection layer, ReLU activation function and full connection layer sequence to obtain the eleventh intermediate normal light reflectance feature map X 2_11 , X' 2_10 Input to the structure composed of a fully connected layer, a ReLU activation function and a fully connected layer in sequence to obtain the eleventh intermediate low light reflectance feature map X' 2_11 ,Will <X 2_11 ,X′ 2_11 >Activate through the Sigmoid activation function to obtain the twelfth intermediate reflectivity feature map binary <X 2_12 ,X' 2_12 >, where X 2_11 The twelfth intermediate normal illumination reflectance feature map X is obtained by activating the Sigmoid function 2_12 , X' 2_11 The twelfth intermediate low light reflectance feature map X' is obtained by activating the Sigmoid function 2_12 ,Will <X 2_9 ,X' 2_9 >with <X 2_12 ,X' 2_12 >Perform element-by-element multiplication to obtain the thirteenth intermediate reflectivity feature map binary <X 2_13 ,X' 2_13 >, where X 2_9 With X 2_12 Perform element-by-element multiplication to obtain the thirteenth intermediate normal light reflectance feature map X 2_13 , X′ 2_9 With X' 2_12 Perform element-by-element multiplication to obtain the thirteenth intermediate low-light reflectance feature map X' 2_13 ,Will <X 2_13 ,X' 2_13 > Input into the structure consisting of a fully connected layer, a ReLU activation function and a fully connected layer sequence to perform feature transformation operation, and obtain the fourteenth intermediate reflectivity feature map binary <X 2_14 ,X' 2_14 >, where X 2_13 Input to the structure composed of the full connection layer, ReLU activation function and full connection layer sequence to obtain the fourteenth intermediate normal light reflectance feature map X 2_14 , X' 2_13 Input to the structure composed of a fully connected layer, a ReLU activation function and a fully connected layer in sequence to obtain the fourteenth intermediate low light reflectance feature map X' 2_14 ; S338: Yes <X 2_14 ,X' 2_14 >After the DropPath operation, the fifteenth intermediate reflectivity feature map binary is obtained <X 2_15 ,X' 2_15 >, where X 2_14 Perform DropPath operation to obtain the fifteenth intermediate normal light reflectance feature map X 2_15 , X' 2_14 Perform DropPath operation to obtain the fifteenth intermediate low light reflectance feature map X' 2_15 ,Will <X 2_15 ,X' 2_15 >with <X 2_8 ,X' 2_8 >Perform the Concat operation to obtain the final second road damage reflectivity feature map binary<R2,R'2> , where X 2_15 With X 2_8 Perform Concat operation to obtain the second normal light road damage reflectivity feature map R2, and convert X' 2_15 With X' 2_8 Perform the Concat operation to obtain the second low-light road damage reflectivity feature map R'2.

7. The method for detecting road damage in low-light substations based on deep learning according to claim 6, characterized in that: The specific steps of S4 are as follows: S41: During the training process, the model extracts data from the training set according to the set batch size for forward calculation to generate prediction results. Subsequently, the model performs backpropagation through the overall loss function L2R-LOSS to optimize the parameters of the substation road damage detection model in low-light environments. S42: After each round of training, the validation set data is used to evaluate the model to determine the generalization ability of the model; S43: To prevent overfitting of the model during training, an early stopping mechanism is introduced during training. If the loss on the validation set does not improve significantly over multiple consecutive training rounds, training will be terminated early. This helps save computing resources and ensures the stability of the model in practical applications. S44: When the training reaches the predetermined number of rounds and the validation set loss is stable or reaches an ideal level, the training process ends, and an optimized low-light substation road damage detection model is finally obtained. The model can accurately detect substation road damage under low-light conditions and maintain good performance in different environments; S45: After training, the model will have appropriate parameter configuration and strong generalization ability, and can be applied to the new substation road damage detection task.

8. The method for detecting road damage in low-light substations based on deep learning according to claim 7, characterized in that: The loss function L2R-LOSS in S41 is composed of the reflectivity learning loss function L r , Reorganization consistency loss function L rc And the target detection loss function L target The weighted composition, its function expression is as follows: L L2R-LOSS =λ g L r +λ f L target +λ q L rc Among them, λ g ,λ f and λ q is the loss weight, the sum of the three is 1; Among them, L r The function expression is as follows: Among them, R'1 is the first low-light road damage reflectivity feature map; R'2 is the first low-light road damage reflectivity feature map; N is the total number of feature points of the R'1 and R'2 feature maps, R' 1_i and R' 2_i Respectively represent the feature response value corresponding to the i-th position in the feature map; and denote the average characteristic response values of R'1 and R'2, respectively, and are the variances of the characteristic responses R'1 and R'2, is the covariance of the characteristic response values between R'1 and R'2, and C1 and C2 are constants to avoid the denominator being zero; Among them, L target The function expression is as follows: Among them, α and β represent the weights of position regression loss and classification loss respectively; x i and x' i Represents the center point of the real box and the predicted box of the i-th target respectively; p i and p' i They represent the true category label and predicted category probability of the i-th target, respectively, and Smooth(x) represents the error between the predicted value and the true value, that is, L rc The function expression is as follows: L rc =||R1-R2||1+||R1-R2||1 Among them, R1 is the first normal light road damage reflectivity characteristic map; R'2 is the second low light road damage reflectivity characteristic map; R'1 is the first low light road damage reflectivity characteristic map; R2 is the second normal light road damage reflectivity characteristic map.

Images