Fine bridge fault detection and reconstruction method based on deep neural radiation field

Abstract

The invention relates to a refined bridge fault detection and reconstruction method based on a deep neural radiation field, and belongs to the field of building artificial intelligence. The method comprises the following steps: obtaining image data and camera pose data of a bridge in a healthy state, constructing a 3D-Brith data set, and preprocessing the 3D-Brith data set to obtain a bridge reference image; constructing a refined bridge fault detection and reconstruction model based on the deep neural radiation field; the reference image and the to-be-detected image enter a fault rate calculation function to calculate the fault rate after passing through a depth image analysis module of the model; a radiation field is generated in a full-connection network of the bridge three-dimensional point pose input model, and a bridge depth image is reconstructed in a volume rendering layer of the model in combination with the camera pose; optimizing by using a loss function to obtain a trained model; and inputting the radiation field and the camera pose visual angle needing to be reconstructed into the volume rendering layer of the trained model to obtain a depth image needing to be reconstructed. According to the invention, defects such as tiny cracks and spalling on the bridge surface can be efficiently identified.

Description

Technical Field

[0001] The present invention belongs to the field of architectural artificial intelligence, and specifically relates to a refined bridge fault detection and reconstruction method based on deep neural radiation field. Background Art

[0002] With the increase in the number of bridges and the extension of their service life, the safety and durability of bridge structures have become increasingly prominent. Bridge fault detection is an important part of ensuring bridge safety and extending its service life. However, traditional detection methods mainly rely on manual inspections and structural health monitoring systems, both of which have significant limitations. Although manual inspections are widely used, they require a lot of manpower and are affected by factors such as the environment and weather, and are inefficient. At the same time, the detection results are highly subjective and are easily limited by the experience level of the inspectors, making it difficult to achieve refined and efficient fault identification. The structural health monitoring system monitors the structural parameters of the bridge, such as strain and vibration, in real time by installing sensors to determine the health status of the bridge. Despite this, the system is costly and the data processing is complex. It is difficult to accurately locate minor damage based on structural parameters alone, especially appearance defects such as cracks or peeling. In addition, image processing and computer vision technologies are widely used in bridge surface defect detection, but due to the complexity of the detection environment, different lighting conditions, changes in viewing angles, and noise interference will affect the accuracy of traditional image processing methods. These methods often perform poorly in extracting defect features such as small cracks or peeling, and their processing efficiency cannot meet real-time requirements when faced with large-scale inspection tasks.

[0003] In order to solve the above problems, image recognition technology based on deep learning has gradually been introduced into the field of bridge fault detection. Deep neural networks can automatically extract multi-level features and achieve remarkable results in classification and recognition tasks. However, the refinement of bridge fault detection requires higher precision and more accurate defect detection capabilities. In this context, the neural radiation field method has attracted much attention because it can construct 3D scenes from a limited number of images and is suitable for 3D reconstruction in complex scenes. The refined deep neural radiation field technology can synthesize a high-resolution 3D depth field with ultra-high sharpness and high-frequency details and compare it with the standard depth map of the bridge, providing accurate positioning and detection capabilities for subtle defects, opening up a new direction for bridge fault detection, and is expected to make breakthrough progress in refined detection and efficient processing. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a refined bridge fault detection and reconstruction method based on deep neural radiation field.

[0005] In order to achieve the above object, the present invention is implemented through the following technical solutions: The present invention provides a refined bridge fault detection and reconstruction method based on deep neural radiation field, comprising the following steps: S1. Obtain bridge image data and corresponding camera pose data taken by a camera in a healthy state of the bridge, construct a 3D-Brith dataset, and preprocess the bridge image data to obtain a bridge benchmark image; S2. Construct a refined bridge fault detection and reconstruction model based on deep neural radiation field, the model includes a deep image analysis module, a failure rate calculation function, a fully connected network, and a volume rendering layer; S3. The reference image of the bridge is obtained by the depth image analysis module to obtain the reference depth image of the bridge, and the image of the bridge to be detected is obtained by the depth image analysis module to obtain the bridge detection depth image; S4. Inputting the bridge reference depth image and the bridge detection depth image into the failure rate calculation function to obtain the failure rate; S5. Input the three-dimensional point pose of the bridge into the fully connected network to obtain the radiation field, and then input the radiation field and the camera pose into the volume rendering layer to obtain the deep reconstruction image of the bridge, and use the loss function to train the model to obtain a trained model; S6. Input the radiation field E and the camera pose and perspective to be reconstructed into the volume rendering layer of the trained model to obtain the depth image to be reconstructed.

[0006] Further, step S1 includes: The corresponding camera pose data refers to the three-dimensional coordinate position of the camera that captured the image in the spatial coordinate system; The image data in the 3D-Brith dataset is preprocessed using the Reinhard global tone mapping algorithm to obtain the bridge benchmark image. .

[0007] Furthermore, the depth image analysis module in step S3 includes a depth encoder, a depth upsampler, and a depth decoder, specifically: The deep encoder includes an image segmentation layer, a patch embedding layer, a position encoding layer, and a patch encoder; a bridge reference image The deep encoder obtains the deep encoding features ; The depth upsampler includes a large-scale merging layer, a large-scale upsampling layer, a medium-scale merging layer, a medium-scale upsampling layer, a small-scale merging layer, a small-scale upsampling layer and a dimension matching layer; the depth encoding feature The deep upsampler obtains the deep upsampled features ; The deep decoder includes a first transposed convolution layer, a ReLU activation function, a second transposed convolution layer, and a sigmoid activation function; the deep upsampling feature Bridge the reference depth image through the depth decoder ; Bridge image to be inspected The depth image of the bridge detection is obtained through the depth image analysis module .

[0008] Furthermore, in step S4, the bridge reference depth image and bridge detection depth image Input into the failure rate calculation function to obtain the failure rate R, which is: S41. The failure rate calculation function sets the following parameters: Bridge benchmark depth image: ; Bridge detection depth image: ; Image height: H; Image width: W; Total number of pixels: ; Number of pixels in the faulty area: ; S42. The fault area is defined as the pixel point whose detection depth is less than the reference depth. The formula is as follows: , Among them, δ represents the indicator function, which is 1 if the conditions in δ are met, otherwise it is 0; S43. Design weighting factor w(x, y), the formula is as follows: , Among them, the max() function means taking the maximum value from the input data; S44. The formula of the failure rate calculation function is as follows: , The calculation result of the failure rate R is a floating point type.

[0009] Furthermore, in step S5, the three-dimensional point pose of the bridge is input into the fully connected network to obtain the radiation field, specifically: The three-dimensional point pose of the bridge , where (x, y, z) are three-dimensional coordinates, calibrated by manual measurement; The fully connected network includes a first hidden layer, a second hidden layer, and a third hidden layer; the first hidden layer contains 24 neurons, and the activation function is the ReLU activation function; the second hidden layer contains 64 neurons, and the activation function is the ReLU activation function; the third hidden layer contains 4 neurons, and the activation function is the ReLU activation function; the three-dimensional point pose of the bridge Input into the fully connected network and get the radiation field E= , σ is the spatial density.

[0010] Furthermore, in step S5, the radiation field and the camera pose are input into the volume rendering layer to obtain a deep reconstruction image of the bridge, specifically: The camera pose =( , ),in is the rotation matrix, is the translation vector, , Calibrated by camera parameters; The volume rendering layer includes a sample point generation layer, a radiation field sampling layer, and a volume integral; in the sample point generation layer, the camera pose Generate sample points in 3D space , the coordinates of each sample point are ; In the radiation field sampling layer, for each sample point , sample according to the radiation field E, and get each sample point The corresponding color value and density ; The volume integral uses the forward integration method to perform weighted summation of the color and perspective of the sample points to obtain the depth reconstruction image of the bridge , the formula is as follows: , Where N represents the number of sample points. represents the perspective transfer function from the camera to the sample point, Represents the distance between adjacent sample points.

[0011] Further, in step S5, the bridge detection depth image is calculated using the Adam optimizer through the MSE loss function. and bridge depth reconstruction images loss.

[0012] The advantages of the present invention are: The present invention proposes a refined bridge fault detection and reconstruction method based on deep neural radiation field, which combines modern deep learning and image processing technology and has significant innovations and beneficial effects. First, the method utilizes the multi-level feature extraction capability of deep neural networks to efficiently identify defects such as tiny cracks and peeling on the surface of the bridge, thereby reducing the impact of human factors and improving the accuracy of detection. At the same time, relying on neural radiation field technology, the method can synthesize high-resolution 3D depth fields, realize multi-perspective positioning of subtle defects, and meet the refined requirements of bridge fault detection. A depth image is generated through an image, and the fault rate is detected through the depth image. If the fault rate is too high, a depth image that is not from the camera perspective can be reconstructed through the generated depth image. By observing the depth image, the stability and reliability of the bridge can be evaluated more systematically.

[0013] In addition, this method can effectively cope with detection challenges in complex environments, overcome the limitations of traditional image processing methods under different lighting conditions and viewing angle changes, and open up a new direction for bridge fault detection. This innovation not only promotes the advancement of intelligent detection technology, but also provides broad application prospects for the development of technology in related fields, with practical value and wide industry application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0015] Figure 1 is a flow chart of the steps of the method of the present invention; Figure 2 This is the depth image of the bridge reconstructed by the method of the present invention. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] Example 1 In this embodiment, Figure 1 As shown, the present invention provides a refined bridge fault detection and reconstruction method based on deep neural radiation field, and the specific steps include: a. Collect camera image data and corresponding camera pose data of a bridge in a healthy state to construct a 3D-Brith dataset; use the Reinhard global tone mapping algorithm to preprocess the image data in the 3D-Brith dataset to obtain the bridge benchmark image ; b. Construct a refined bridge fault detection and reconstruction model based on deep neural radiation field, the model includes a deep image analysis module, a fault rate calculation function, a fully connected network, and a volume rendering layer; the deep image analysis module includes a deep encoder, a deep upsampler, and a deep decoder; c. Bridge benchmark image Input into the deep encoder to obtain deep encoding features ; d. Deep encoding features Input into the depth upsampler to obtain the depth upsampled features ; e. Upsample the depth feature Input into the depth decoder to obtain the bridge benchmark depth image ; f. The bridge image to be detected Input into the depth encoder, depth upsampler, and depth decoder in step cd in sequence to obtain the bridge detection depth image ; g. Bridge benchmark depth image and bridge detection depth image Input into the failure rate calculation function to obtain the failure rate R; h. The three-dimensional point pose Input into the fully connected network to obtain the radiation field E; i. Combine the radiation field E and the camera pose Input into the volume rendering layer to obtain the depth reconstruction image of the bridge ; j. Calculate the bridge detection depth image using the Adam optimizer through the MSE loss function and bridge depth reconstruction images The loss is , and the parameters of the model in iterative steps h) to i) are obtained to obtain the trained model , the model M includes a depth encoder, a depth upsampler, a depth decoder, a failure rate calculation function, a fully connected network, and a volume rendering layer; k. Combine the radiation field E and the pose view to be reconstructed Input into the volume rendering layer trained in step i), and output the depth image to be reconstructed The pose and perspective to be reconstructed The perspective that needs to be reconstructed is outside the camera perspective.

[0018] Specifically, step a is: The 3D-Brith dataset uses 6 sets of fixed-point cameras to collect image data from a bridge in a healthy state. The corresponding camera pose data is measured based on the position and attitude of the camera. The corresponding camera pose data refers to the three-dimensional coordinate position of the camera in the spatial coordinate system. ; Create 6 folders to store the 6 groups of image data and the corresponding camera pose data, named: 1, 2, 3, 4, 5, 6; representing 6 different camera positions; this article uses the first group of point data for illustration, and the data operation methods for the remaining points are the same as the first group.

[0019] Specifically, step c includes the following steps: c-1) The deep encoder consists of an image segmentation layer, a patch embedding layer, a position encoding layer, and a patch encoder in sequence.

[0020] c-2) The image segmentation layer first divides the input image into several fixed-size blocks, each of which is called a patch, and the edges of each patch overlap each other. The input image size is set to 225×225, the patch size is set to 16×16, and the overlap size is 1. This means that each image will be divided into 225 patches.

[0021] c-3) The patch embedding layer expands each patch and linearly projects it into a high-dimensional space to form a fixed-length patch embedding vector. This paper uses a convolutional layer to implement patch expansion, where the convolution kernel size is 3*3, the number of convolution kernels is 8, the stride is 1, and 'same' padding is used. In addition, a fully connected layer can also be used for patch expansion.

[0022] c-4) The position encoding layer uses the iRPE method to add relative position encoding to each patch embedding vector This method can better handle transformation invariance and is more effective for the deep vision task in this paper.

[0023] c-5) The patch encoder consists of a first convolutional layer, a second convolutional layer, a ReLU activation function, and a pooling layer. The convolution kernel size of the first convolutional layer is 1*1, the number of convolution kernels is 8, the stride is 1, and 'same' padding is used. The convolution kernel size of the second convolutional layer is 3*3, the number of convolution kernels is 16, the stride is 1, and 'same' padding is used. The second pooling layer uses maximum pooling, the pooling window size is 3*3, and the stride is 1.

[0024] c-6) Bridge benchmark image Input into the image segmentation layer, patch embedding layer, position encoding layer, and patch encoder to obtain deep encoding features .

[0025] Specifically, step d comprises the following steps: d-1) The depth upsampler consists of a large-scale merging layer, a large-scale upsampling layer, a medium-scale merging layer, a medium-scale upsampling layer, a small-scale merging layer, a small-scale upsampling layer, and a dimension matching layer.

[0026] d-2) The large-scale merging layer encodes features according to depth The position encoding selects the 1st to 160th patches for merging. In order, 16 patches are selected as a group, and each group is arranged from left to right and from top to bottom to form a 64×64 feature map. After arrangement, a total of 10 groups of 64×64 feature maps are formed. The 10 groups of feature maps are added and averaged to obtain a 64×64 large-scale merged feature.

[0027] d-3) The large-scale upsampling layer uses a bilinear interpolation method to expand the dimension of the large-scale merged features to 512×512.

[0028] d-4) The mesoscale merging layer encodes features according to depth The position encoding selects the 161-209th patches for merging. In order, 4 patches are selected as a group, and each group is arranged from left to right and from top to bottom to form a 32×32 feature map. After arrangement, a total of 12 groups of 32×32 feature maps are formed. The 12 groups of feature maps are added and averaged to obtain a 32×32 mesoscale merged feature.

[0029] d-5) The mid-scale upsampling layer uses a bilinear interpolation method to expand the dimension of the mid-scale merged features to 256×256.

[0030] d-6) The small-scale merging layer encodes features according to depth The position encoding selects the 210th to 225th patches for merging. The 16 patches are added and averaged to obtain a 16×16 small-scale merged feature.

[0031] d-7) The small-scale upsampling layer uses a bilinear interpolation method to expand the dimension of the small-scale merged features to 64×64.

[0032] d-8) The dimension matching layer adds and fuses the large-scale merged features, medium-scale merged features, and small-scale merged features expanded from d-2), d-4), and d-6) to obtain the deep upsampling features. .

[0033] Specifically, step e includes the following steps: e-1) The deep decoder is composed of a first transposed convolutional layer, a ReLU activation function, a second transposed convolutional layer, and a sigmoid activation function. The convolution kernel size of the first transposed convolutional layer is 3*3, the number of convolution kernels is 256, the stride is 1, and the 'same' padding is used. The convolution kernel size of the second transposed convolutional layer is 3*3, the number of convolution kernels is 512, the stride is 1, and the 'same' padding is used.

[0034] e-2) Upsampling the depth feature Input into the first transposed convolution layer, ReLU activation function, second transposed convolution layer, sigmoid activation function to obtain the bridge benchmark depth image .

[0035] Specifically, step g) comprises the following steps: g-1) The failure rate calculation function sets the following parameters: Bridge benchmark depth image: Bridge detection depth image: Image height: H Image width: W Total number of pixels: Number of pixels in the faulty area: g-2) The fault area is defined as those pixels whose detection depth is less than the reference depth, so it can be expressed as: Where δ represents the indicator function, which satisfies: g-3) A weighting factor (w(x, y)) is introduced to consider the importance of different regions. The max() function takes the maximum value from the input data: g-4) The calculation process of the failure rate calculation function is as follows: The calculation result of the failure rate (R) is of floating point type and is output as a percentage.

[0036] Specifically, step h comprises the following steps: h-1) The three-dimensional point pose , where (x, y, z) are three-dimensional coordinates calibrated by manual measurement.

[0037] h-2) The fully connected network consists of a first hidden layer, a second hidden layer, and a third hidden layer. The first hidden layer contains 24 neurons, and the activation function is ReLU. The second hidden layer contains 64 neurons, and the activation function is ReLU. The third hidden layer contains 4 neurons, and the activation function is ReLU.

[0038] h-3) The three-dimensional point pose Input into the fully connected network and get the radiation field E= , σ is the spatial density.

[0039] Specifically, step i includes the following steps: i-1) The camera pose =( , ),in is the rotation matrix, is the translation vector, which is calibrated by the camera parameters.

[0040] i-2) The volume rendering layer is composed of a sample point generation layer, a radiation field sampling layer, and a volume integral. The sample point generation layer starts from the camera viewpoint and is based on the camera pose. Generate a series of sample points in 3D space , the coordinates of each sample point are The radiation field sampling layer is used for each sample point , sample according to the radiation field E and get the corresponding color value and density The volume integral uses the forward integration method to perform a weighted summation of the color and perspective of the sample points. The specific formula is as follows: Where N is the number of sample points, is the perspective transfer function from the camera to the sample point, defined as: here is the distance between adjacent sample points.

[0041] i-3) The radiation field E and the camera pose Input into the volume rendering layer to obtain the depth reconstruction image of the bridge ; Example 2 In practical applications, the refined bridge fault detection and reconstruction method based on deep neural radiation field of the present invention has been experimentally verified on multiple bridge structures. Several different types of bridges were selected for the experiment, including highway bridges, railway bridges and pedestrian bridges. Each bridge was photographed under different environmental conditions, covering direct sunlight, rainy weather and low illumination at night.

[0042] In order to demonstrate the advantages of the refined bridge fault detection and reconstruction method based on deep neural radiation field in the present invention over traditional detection technology and advanced detection technology Mentor DFT, the following are the comparative experimental results of the two on multiple key performance indicators. The data table shows indicators such as recognition rate, processing efficiency and positioning accuracy.

[0043] Table 1 Experimental comparison between the method of the present invention and the existing method As can be seen from Table 1, the refined bridge fault detection and reconstruction method based on deep neural radiation field proposed in the present invention shows significant advantages in various performance indicators. The microcrack recognition rate of the method of the present invention reached 92%, which is 25 percentage points higher than the traditional visual inspection method and 11 percentage points higher than the advanced method Mentor DFT. The recognition rate of spalling defects also reached 90%, which is 30 percentage points higher than the traditional method and 15 percentage points higher than the advanced method Mentor DFT. This shows that the present invention can more effectively discover potential problems in bridges and provide more reliable data support for preventive maintenance. In terms of processing efficiency, the method of the present invention only takes 30 minutes to complete the processing of 100 images, while the traditional method takes 120 minutes and the advanced method Mentor DFT takes 90 minutes. This shows the superiority of the present invention in real-time detection and emergency response capabilities, making it more feasible in practical applications. For the accuracy of defect positioning, the method of the present invention is controlled within 5 mm, which is significantly lower than the traditional visual inspection and advanced method Mentor DFT (10-15 mm and 8-12 mm, respectively), reflecting the outstanding performance of the present invention in accuracy.

[0044] Through the above comparison data, it can be clearly seen that the bridge fault detection method based on deep neural radiation field of the present invention has significant advantages in key performance indicators such as recognition rate, processing efficiency and positioning accuracy. These advantages not only improve the effectiveness and practicality of the detection method, but also provide important guarantees for bridge safety management and help improve the health monitoring level of the overall structure.

[0045] exist Figure 2In the figure, we can see that the generated depth image clearly depicts the defects on the bridge surface at high resolution. The red areas represent serious defects, which greatly facilitates subsequent maintenance decisions. The marked defect points are highly consistent with the actual detection results, further verifying the accuracy and practicality of this method.

[0046] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A refined bridge fault detection and reconstruction method based on deep neural radiation field, characterized in that: The following steps are involved: S1. Obtain bridge image data and corresponding camera pose data taken by a camera in a healthy state of the bridge, construct a 3D-Brith dataset, and preprocess the bridge image data to obtain a bridge benchmark image; S2. Construct a refined bridge fault detection and reconstruction model based on deep neural radiation field, the model includes a deep image analysis module, a failure rate calculation function, a fully connected network, and a volume rendering layer; S3. The reference image of the bridge is passed through the depth image analysis module to obtain the reference depth image of the bridge, and the image of the bridge to be detected is passed through the depth image analysis module to obtain the bridge detection depth image; S4. Inputting the bridge reference depth image and the bridge detection depth image into the failure rate calculation function to obtain the failure rate; S5. Input the three-dimensional point pose of the bridge into the fully connected network to obtain the radiation field, and then input the radiation field and the camera pose into the volume rendering layer to obtain the deep reconstruction image of the bridge, and use the loss function to train the model to obtain a trained model; S6. Input the radiation field and the camera pose and perspective to be reconstructed into the volume rendering layer of the trained model to obtain the depth image to be reconstructed.

2. The refined bridge fault detection and reconstruction method based on deep neural radiation field according to claim 1 is characterized in that: Step S1 includes: The corresponding camera pose data refers to the three-dimensional coordinate position of the camera that captured the image in the spatial coordinate system; The image data in the 3D-Brith dataset is preprocessed using the Reinhard global tone mapping algorithm to obtain the bridge benchmark image. .

3. The refined bridge fault detection and reconstruction method based on deep neural radiation field according to claim 2 is characterized in that: The depth image analysis module in step S3 includes a depth encoder, a depth upsampler, and a depth decoder, specifically: The deep encoder includes an image segmentation layer, a patch embedding layer, a position encoding layer, and a patch encoder; a bridge reference image The deep encoder obtains the deep encoding features ; The depth upsampler includes a large-scale merging layer, a large-scale upsampling layer, a medium-scale merging layer, a medium-scale upsampling layer, a small-scale merging layer, a small-scale upsampling layer and a dimension matching layer; the depth encoding feature The deep upsampler obtains the deep upsampled features ; The deep decoder includes a first transposed convolution layer, a ReLU activation function, a second transposed convolution layer, and a sigmoid activation function; the deep upsampling feature Bridge the reference depth image through the depth decoder ; Image of the bridge to be inspected The depth image of the bridge detection is obtained through the depth image analysis module .

4. The refined bridge fault detection and reconstruction method based on deep neural radiation field according to claim 3 is characterized in that: In step S4, the bridge reference depth image and bridge detection depth image Input into the failure rate calculation function to obtain the failure rate R, which is: S41. The failure rate calculation function sets the following parameters: Bridge benchmark depth image: ; Bridge detection depth image: ; Image height: H; Image width: W; Total number of pixels: ; Number of pixels in the faulty area: ; S42. The fault area is defined as the pixel point whose detection depth is less than the reference depth. The formula is as follows: ， Among them, δ represents the indicator function, which is 1 if the conditions in δ are met, otherwise it is 0; S43. Design weighting factor w(x, y), the formula is as follows: ， Among them, the max() function means taking the maximum value from the input data; S44. The formula of the failure rate calculation function is as follows: ， The calculation result of the failure rate R is a floating point type.

5. The refined bridge fault detection and reconstruction method based on deep neural radiation field according to claim 4 is characterized in that: In step S5, the three-dimensional point pose of the bridge is input into the fully connected network to obtain the radiation field, specifically: The three-dimensional point pose of the bridge , where (x, y, z) are three-dimensional coordinates, calibrated by manual measurement; The fully connected network includes a first hidden layer, a second hidden layer, and a third hidden layer; the first hidden layer contains 24 neurons, and the activation function is the ReLU activation function; the second hidden layer contains 64 neurons, and the activation function is the ReLU activation function; the third hidden layer contains 4 neurons, and the activation function is the ReLU activation function; the three-dimensional point pose of the bridge Input into the fully connected network and get the radiation field E= , σ is the spatial density.

6. The refined bridge fault detection and reconstruction method based on deep neural radiation field according to claim 5 is characterized in that: In step S5, the radiation field and camera pose are input into the volume rendering layer to obtain a deep reconstruction image of the bridge, specifically: The camera pose =( , ),in is the rotation matrix, is the translation vector, , Calibrated by camera parameters; The volume rendering layer includes a sample point generation layer, a radiation field sampling layer, and a volume integral; in the sample point generation layer, the camera pose Generate sample points in 3D space , the coordinates of each sample point are ; In the radiation field sampling layer, for each sample point , sample according to the radiation field E, and get each sample point The corresponding color value and density ; The volume integral uses the forward integration method to perform weighted summation of the color and perspective of the sample points to obtain the depth reconstruction image of the bridge , the formula is as follows: ， Where N represents the number of sample points. represents the perspective transfer function from the camera to the sample point, Represents the distance between adjacent sample points.

7. The refined bridge fault detection and reconstruction method based on deep neural radiation field according to claim 6 is characterized in that: In step S5, the bridge detection depth image is calculated using the Adam optimizer through the MSE loss function and bridge depth reconstruction images loss.

Images