A method for detecting defects of a tunnel chute based on an FDEResUNet network

Abstract

The application provides a method for detecting defects of a tunnel chute based on an FDEResUNet network, and the method comprises the following steps: collecting a tunnel chute image to be detected; inputting the tunnel chute image to be detected into a pre-trained chute defect detection model; the chute defect detection model is constructed based on an improved FDEResUNet network; and outputting a target tunnel chute image with a defective target class position corresponding to the tunnel chute image to be detected by using the chute defect detection model. Through the automatic image processing technology, the need for manual inspection is reduced, and the detection efficiency is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of railway construction inspection technology, and in particular to a method for detecting tunnel slippage defects based on the FDEResUNet network. Background Technology

[0002] In recent years, railways have developed rapidly, with an increasing number of construction and operation mileages. As a result, there is a growing need for standardized and safety inspections of railway lines during the construction and operation of railway tunnels.

[0003] Slippage in high-speed railway tunnels plays a crucial role in the integrity and stable operation of railway tunnel construction. If slippage defects are not detected and addressed promptly in the early stages, it will adversely affect the smooth operation of existing power transmission lines. Due to the complex environment of railway construction sites, manual inspection suffers from low efficiency, low measurement accuracy, and blind spots, which can lead to safety risks.

[0004] The tunnel environment is dimly lit and spans a long distance. The surface color of the tunnel chutes is dark and the boundaries are blurred. The acquired image data is large in volume, and there are also images with poor lighting and blurriness. However, labeling and annotating the images of railway tunnel chutes requires a long period of manual measurement, wasting a lot of time for professional technicians, which increases the difficulty of identifying and distinguishing defective chutes. Through data analysis, it was found that defects are frequently observed in the chutes and nearby important areas in the acquired chute images, and the defects are not limited to fixed locations but can appear all around the chutes. Images depicting chute defects usually show defect characteristics in multiple areas. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide a method for detecting tunnel slippage defects based on the FDEResUNet network.

[0006] This invention provides a method for detecting tunnel slippage defects based on the FDEResUNet network, the method comprising:

[0007] Acquire images of the tunnel trough to be inspected;

[0008] The image of the tunnel chute to be detected is input into a pre-trained chute defect detection model; the chute defect detection model is built based on an improved FDEResUNet network.

[0009] The trough defect detection model is used to output the target tunnel trough image with the defect target category location corresponding to the tunnel trough image to be detected.

[0010] Optionally, the training process of the groove defect detection model includes:

[0011] Collect sample images of tunnel chutes, generate a sample dataset based on the sample images, and segment the sample dataset to obtain a training set and a test set;

[0012] A groove defect detection model based on the improved FDEResUNet network is constructed, and the sample images in the training set are input into the groove defect detection model to train the groove defect detection model.

[0013] The tunnel chute images in the test set are input into the trained chute defect detection model to output the chute defect target segmentation result.

[0014] Optionally, the groove defect detection model includes:

[0015] Encoder layer: The input tunnel chute image to be detected first passes through an encoder layer consisting of four ResNetBlock downsampling residual blocks. Each downsampling residual block R extracts hierarchical features from the input tunnel chute image to be detected. The output Ii of each downsampling residual block R is used as the input to the frequency domain module FDB. A portion of Li is added to a skip connection, and the other portion Hi is used as the input to the next downsampling residual block R. Finally, the encoder output H4 is input to the relay block.

[0016] Intermediate Block: The intermediate block contains 1024 feature channels and is used to connect the encoder layer and the decoder layer;

[0017] Decoder layer: The data from the intermediate block is input into the first upsampled convolutional block C of the decoder layer. The decoder stage contains four upsampled convolutional blocks C. The skip connection takes the Li output from each downsampled residual block R in the encoder after passing through the frequency domain module FDB and the output Ji from the previous encoder layer as input, and the output result is Ji+1;

[0018] Dual-layer enhancement block module (DLE): The output J4 of the last upsampling module in the decoder layer is used as the input of the dual-layer enhancement block module (DLE). After passing through two dual-layer enhancement blocks (DLE), the final enhancement result K is output as the final output of the model.

[0019] Optionally, the dual-layer enhancement module has two parallel layers, each layer containing a two-dimensional convolution Conv, a Dropout value of 0.1, batch normalization (BN), and a ReLU activation function.

[0020] Optionally, the downsampled residual block R contains a two-dimensional convolution Conv, batch normalization BN, Dropout with a value of 0.3, and a ReLU activation function, doubling the number of feature channels.

[0021] The present invention also provides an apparatus for detecting tunnel slippage defects based on an FDEResUNet network. The apparatus includes a processor and a memory; the memory is used to store program code and transmit the program code to the processor; the processor is used to execute the method described above according to the instructions in the program code.

[0022] The present invention also provides a computer-readable storage medium for storing program code for performing the method described in any of the preceding embodiments.

[0023] The present invention also provides a computing device, which includes the aforementioned apparatus for detecting tunnel slippage defects based on the FDEResUNet network.

[0024] This invention presents a method for detecting tunnel chute defects based on the FDEResUNet network. The method employs the FDEResUNet network for image detection. Firstly, it uses an improved wavelet packet transform module and a double-layer enhancement block to optimize the single skip link function in the original UNet structure, thereby enhancing the frequency information present in the chute image and identifying the boundary of the defect target. Secondly, it processes the acquired tunnel chute images based on the improved FDEResUNet network. This model takes the original chute image as input, passes through an encoder layer, a decoder layer, and a double-layer enhancement block, and finally obtains the chute defect target segmentation result.

[0025] The above and other objects, advantages and features of the present invention will become more apparent to those skilled in the art from the following detailed description of specific embodiments of the invention in conjunction with the accompanying drawings. Attached Figure Description

[0026] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0027] Figure 1 This is an embodiment of the method for detecting tunnel slippage defects based on the FDEResUNet network in this invention;

[0028] Figure 2 This is a schematic diagram of the training process of the groove defect detection model according to an embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram of the structure of the groove defect detection model according to an embodiment of the present invention. Detailed Implementation

[0030] The embodiments of the present invention are described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the present invention and are not restrictive.

[0031] This invention provides a method for detecting tunnel slippage defects based on the FDEResUNet network, the method comprising:

[0032] S1, Acquire images of the tunnel chute to be detected;

[0033] S2, input the image of the tunnel chute to be detected into a pre-trained chute defect detection model; the chute defect detection model is built based on an improved FDEResUNet network;

[0034] S3, using the chute defect detection model, output the target tunnel chute image with the defect target category location corresponding to the tunnel chute image to be detected.

[0035] In practical applications, chute images can be acquired first from the tunnel, avoiding dark environments as much as possible during the acquisition process; then, the acquired chute images are uniformly input into the FDEResUNet model in 256×256 format for defect detection; finally, the chute images with segmented defect areas are output in 256×256 format.

[0036] In this example, the chute images are categorized into two types: defective and normal. Specific types of defects include chute misalignment, chute deformation, foreign objects in the chute, concrete coverage, missing concrete, uneven cement, damage to the galvanized layer, and burial depth.

[0037] The advantages and beneficial effects of this invention compared to traditional solutions are as follows: By using automated image processing technology, the need for manual inspection is reduced, thus significantly improving detection efficiency; Utilizing deep learning models and automated image processing technology, the need for manual inspection is reduced, thus significantly improving detection efficiency; The automated detection system is not limited by human vision, can cover a wider area, and reduce blind spots caused by complex environments or obstructed vision; Through wavelet packet transform (WPT) and dual-layer enhancement blocks (DLE), the model can more effectively process and analyze image data, especially in environments with poor lighting conditions such as tunnels; By reducing the frequency of manual inspections and improving detection accuracy, it helps to reduce the maintenance costs of railway tunnels in the long run.

[0038] In some alternative embodiments, combined with Figure 2 As shown, the training process for the groove defect detection model can be as follows:

[0039] S11 acquires sample images of the tunnel chute, generates a sample dataset based on the sample images, and segments the sample dataset to obtain a training set and a test set;

[0040] S12, Construct a groove defect detection model based on the improved FDEResUNet network, and input the sample images in the training set into the groove defect detection model to train the groove defect detection model;

[0041] S13, input the tunnel chute images in the test set into the trained chute defect detection model to output the chute defect target segmentation result.

[0042] like Figure 3 As shown, the groove defect detection model structure in this embodiment may include the following modules:

[0043] I. Encoder layer.

[0044] The input image of the tunnel chute to be detected first passes through an encoder layer consisting of four ResNet Block downsampling residual blocks. Each downsampling residual block R extracts hierarchical features from the input image of the tunnel chute to be detected. The output Ii of each downsampling residual block R is used as the input to the frequency domain block (FDB). A portion of Li is added to the skip connection, and another portion Hi is used as the input to the next downsampling residual block R. Finally, the encoder output H4 is input to the relay block.

[0045] The downsampled residual block R contains a two-dimensional convolution Conv, batch normalization BN, Dropout set to 0.3, and ReLU activation function, doubling the number of feature channels.

[0046] II. Frequency Domain Module.

[0047] The frequency domain division module in this embodiment employs wavelet packet transform. Wavelet packet transform (WPT) is a signal analysis method based on wavelet transform, offering more flexible frequency analysis capabilities than traditional wavelet transform, allowing for arbitrary levels of image decomposition. The advantage of wavelet packet transform lies in its ability to not limit the decomposition levels or the number of bands, allowing for further decomposition of all bands in each decomposition step. It is widely used in image processing. Wavelet packet transform can be used to extract image features such as edges and textures. Thorough extraction of these features is crucial for this task.

[0048] To further improve the training effect, preserve basic image information and reduce feature dimensionality, two-dimensional wavelet packet decomposition is adopted. Since tunnel troughs are mostly regular horizontal and vertical stripes, the focus is on the two dimensions of horizontal and vertical, and wavelet packet decomposition is applied emphasizing.

[0049] In Python, wavelet packet transform can be implemented using various third-party libraries, the most famous of which is "PyWavelets", or "pywt" for short. "PyWavelets" is an open-source Python library that provides support for various wavelet transforms, including wavelet packet transform (WPT). The WPT function performs a wavelet packet transform on the output image features Ii of each downsampled residual block. The output of WPT is a waveletPacket2D object, which stores information about all possible wavelet packet decomposition subbands in the image.

[0050] Each decomposition step generates more sub-bands. In the first-level decomposition, the original image is divided into four sub-bands: LL (low frequency), LH (horizontal high frequency), HL (vertical high frequency), and HH (diagonal high frequency), with the HH sub-band actively removed to avoid noise interference. In the second-level decomposition, the LL (low frequency) sub-band is no longer decomposed, treating LL as low-frequency information. The LH (horizontal high frequency) sub-band is selectively further decomposed to obtain LH_LL (lower-frequency horizontal high frequency), LH_LH (second-highest horizontal frequency), and LH_HL (second-highest vertical frequency). The HL (vertical high frequency) sub-band is further decomposed to obtain HL_LL (lower-frequency vertical high frequency), HL_LH (second-highest horizontal frequency), and HL_HL (second-highest vertical frequency).

[0051] Ultimately, we choose to unify the four subbands LH_LH, LH_HL, HL_LH, and HL_HL as the input Hi (i=1, 2, 3, 4) of the next residual block; and unify the three subbands LL, LH_LL, and HL_LL as Li. After passing through a skip connection and adding the previous upsampled feature map Ji element-wise with the decoder, we recover the output feature map Ji+1. (In image processing, approximation coefficients are often used to represent the coarse outline of an image. In the PyWavelets library, approximation coefficients can be obtained through WPT. Approximation coefficients can be obtained by accessing specific subband paths (e.g., 'a'*level, where level refers to the level)).

[0052] III. Transfer Block.

[0053] The relay block contains 1024 feature channels and is used to connect the encoder layer and the decoder layer.

[0054] The intermediate block, as a crucial part connecting the encoder and decoder layers, is an essential component of the UNet network structure, enhancing its feature extraction capabilities. The intermediate block in the model contains 1024 feature channels and includes 2D convolution (Conv), batch normalization (BN), and ReLU activation functions.

[0055] IV. Decoder Layer.

[0056] The data from the intermediate block is input into the first upsampled convolutional block C of the decoder layer. The decoder stage contains four upsampled convolutional blocks C. The skip connection takes the Li output from each downsampled residual block R in the encoder after passing through the frequency domain module FDB and the output Ji from the previous encoder layer as input, and the output result is Ji+1.

[0057] The decoder integrates the encoder's feature information through multiple upsampling blocks to gradually recover the size of the feature map, ultimately outputting the required feature map. The upsampling stage consists of three convolutional blocks with an upsampling stride of 2. Each convolutional block (CB) contains a 2D 3×3 kernel convolution (Conv), batch normalization (BN), and a dropout of 0.2. The number of feature channels is halved (512, 256, 128, 64). Skip connections transmit the spatial information downsampled from each layer of the encoder, after passing through a wavelet packet transform module and a lightweight cross-attention module, to the upsampling decoder layer of each layer.

[0058] V. Dual layer enhancement block (DLE) module

[0059] The output J4 of the last upsampling module in the decoder layer is used as the input of the dual-layer enhancement module DLE. After passing through two dual-layer enhancement modules DLE, the final enhancement result K is output as the final output of the model.

[0060] Due to improper operation during the installation of the tunnel chute, there may be minor defects near the chute. Low-resolution images of the chute were captured in the dimly lit tunnel. To better identify these defects, a dual-layer enhancement module (DLE) was added after the decoder layer. The dual-layer enhancement module has two parallel layers, each containing a 2D convolution (Conv), a dropout of 0.1, batch normalization (BN), and a ReLU activation function. The addition of the dual-layer enhancement module helps the model expand the pixel range and consider more contextual information when analyzing images.

[0061] To accurately segment defective targets in tunnel trough images, a model capable of recognizing local features is needed. The UNet architecture offers flexibility and powerful segmentation capabilities, and variants of UNet have been extensively studied and applied to other types of image segmentation tasks. The UNet architecture is a CNN model that uses skip connections between the encoder and decoder to ensure smooth information flow during backpropagation, facilitating more efficient model training.

[0062] However, traditional connections make depth difficult to predict; excessive depth can lead to small gradients and slow down the training process. To overcome this problem, this invention adds wavelet packet transform (WPT) to the traditional UNet model, that is, it uses a frequency domain module to replace the original direct connection between the encoder and decoder, utilizing the frequency information present in the strong groove image.

[0063] The encoder and decoder in a typical UNet architecture consist of convolutional blocks. We optimize the downsampling stage of the encoder into a ResNet Block (RB), which contains a 2D convolution (Conv), batch normalization (BN), a dropout of 0.3, and a ReLU activation function. The number of feature channels is doubled (64, 128, 256, 512). This design of each downsampling block compensates for the vanishing gradient problem that occurs in the original convolutional blocks as training depth increases, and also accelerates the model's convergence speed. This is crucial for improving the accuracy of target detection.

[0064] Each residual block incorporates a Dropout layer. The downsampling phase is set to 0.3 to prevent training on the same feature set repeatedly. The upsampling phase is set to 0.2, continuously incorporating Dropout to suppress overfitting during model training. Adding Dropout layers increases the randomness of the training data. The ResNet block, combined with Dropout, greatly aids in extracting the boundaries of the chute defect portion during the entire model training process.

[0065] The tunnel chute defect detection method based on the improved FDEResUNet network provided in this invention involves the following steps: For a tunnel chute image to be detected, the encoder layer performs downsampling to generate feature maps Ii (i=1, 2, 3, 4) in the chute image. After each upsampling block, the decoder adds the upsampled feature map element-wise to the feature map from the WPT to recover the output feature map Ji (i=1, 2, 3, 4). The double-layer enhancement block takes the feature map J4 output from the decoder layer as input and finally outputs the enhanced feature map K. Finally, the model output locates the tunnel chute image marked with the defect target category.

[0066] This invention also provides an apparatus for detecting tunnel sluice defects based on an FDEResUNet network. The apparatus includes a processor and a memory. The memory stores program code and transmits the program code to the processor. The processor executes the method for detecting tunnel sluice defects based on an FDEResUNet network according to the instructions in the program code. In this embodiment, the sluice defect detection algorithm is named "FDEResUNet," which involves a Res structure, where F is the frequency domain block and DE is the double-layer enhancement block. These two modules are described in the technical disclosure. UNet is a classic image segmentation network.

[0067] This invention also provides a computer-readable storage medium for storing program code for executing the method for detecting tunnel slippage defects based on the FDEResUNet network described in the above embodiments.

[0068] This invention also provides a computing device, which includes the apparatus for detecting tunnel slippage defects based on the FDEResUNet network described in the above embodiments.

[0069] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting tunnel slippage defects based on FDEResUNet network, characterized in that, The method includes: Acquire images of the tunnel trough to be inspected; The image of the tunnel chute to be detected is input into a pre-trained chute defect detection model; the chute defect detection model is built based on an improved FDEResUNet network. The trough defect detection model is used to output a target tunnel trough image with the location of the defect target category corresponding to the tunnel trough image to be detected; The groove defect detection model includes: an encoder layer, a frequency domain module, a relay block, a decoder layer, and two dual-layer enhancement modules; The encoder layer uses four ResNet Block downsampling residual blocks. Each downsampling residual block R extracts hierarchical features from the input image of the tunnel chute to be detected. The output Ii of each downsampling residual block R is used as the input of the frequency domain module FDB. A part of Li is added to the skip connection, and another part Hi is used as the input of the next downsampling residual block R. Finally, the encoder output H4 is input to the relay block. The frequency domain segmentation module employs wavelet packet transform to perform multi-level frequency decomposition on the output Ii, including: in the first-level decomposition, the feature map is decomposed into four sub-bands: low-frequency LL, horizontal high-frequency LH, vertical high-frequency HL, and diagonal high-frequency HH, and the HH sub-band causing noise interference is actively deleted; in the second-level decomposition, the LL sub-band is not decomposed, and the horizontal high-frequency LH sub-band is selectively decomposed to obtain lower-frequency horizontal high-frequency LH_LL, horizontal second-high-frequency LH_LH, and vertical second-high-frequency LH_HL, and the vertical high-frequency HL sub-band is further decomposed to obtain lower-frequency vertical high-frequency HL_LL, horizontal second-high-frequency HL_LH, and vertical second-high-frequency HL_HL; the four sub-bands LH_LH, LH_HL, HL_LH, and HL_HL are used as the input Hi of the next downsampling residual block, and the three sub-bands LL, LH_LL, and HL_LL are unified as Li and fused with the decoder layer through skip connections; The relay block contains 1024 feature channels for connecting the encoder layer and the decoder layer.

2. The method according to claim 1, characterized in that, The training process for the groove defect detection model includes: Collect sample images of tunnel chutes, generate a sample dataset based on the sample images, and segment the sample dataset to obtain a training set and a test set; A groove defect detection model based on the improved FDEResUNet network is constructed, and the sample images in the training set are input into the groove defect detection model to train the groove defect detection model; The tunnel chute images in the test set are input into the trained chute defect detection model to output the chute defect target segmentation result.

3. The method according to claim 1, characterized in that, The decoder layer: The data from the relay block is input into the first upsampled convolutional block C of the decoder layer. The decoder stage contains four upsampled convolutional blocks C. The skip connection takes the Li output from each downsampled residual block R in the encoder after passing through the frequency domain module FDB and the output Ji from the previous encoder layer as input, and the output result is Ji+1; The dual-layer enhancement module DLE takes the output J4 of the last upsampling module in the decoder layer as the input of the dual-layer enhancement module DLE, and outputs the final enhancement result K after passing through two dual-layer enhancement modules DLE, which is used as the final output of the model.

4. The method according to claim 1, characterized in that, Each layer in the dual-layer enhancement module includes a two-dimensional convolution Conv, Dropout set to 0.1, batch normalization (BN), and ReLU activation function.

5. The method according to claim 3, characterized in that, The relay block contains 1024 feature channels.

6. The method according to claim 1, characterized in that, The downsampled residual block R contains a two-dimensional convolution Conv, batch normalization BN, Dropout with a value of 0.3, and a ReLU activation function, doubling the number of feature channels.

7. A device for detecting tunnel slippage defects based on the FDEResUNet network, characterized in that, The device includes a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is configured to execute the method according to any one of claims 1-6 according to the instructions in the program code.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store program code for performing the method according to any one of claims 1-6.

9. A computing device, characterized in that, The computing device includes the apparatus for detecting tunnel slip defects based on the FDEResUNet network as described in claim 7.

Images