Method for detecting tunnel chute defects based on FDEResUNet network

Abstract

The invention provides a method for detecting tunnel chute defects based on an FDEResUNet network. The method comprises the following steps: acquiring an image of a to-be-detected tunnel chute; inputting the to-be-detected tunnel chute image into a pre-trained chute defect detection model; the chute defect detection model is constructed on the basis of an improved FDEResUNet network; and outputting a target tunnel chute image with a defect target category position corresponding to the to-be-detected tunnel chute image by using the chute defect detection model. According to the invention, through an automatic image processing technology, the requirement of manual inspection is reduced, so that the detection efficiency is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of railway construction inspection, and in particular to a method for detecting tunnel chute defects based on a FDEResUNet network. Background Art

[0002] In recent years, railways have developed rapidly, with more and more mileage under construction and in operation. Along with this, there are more and more inspections on the standardization and safety of railway lines during the construction and operation of railway tunnels.

[0003] The chute of a high-speed railway tunnel plays a very important role in the integrity and stable operation of the railway tunnel construction. If the chute defects cannot be promptly checked in the early stage, it will have an adverse impact on the smooth operation of the existing power supply system of the transmission line. Due to the complex environment of the railway construction site, manual inspection has low operating efficiency, low measurement accuracy, and detection blind spots, which will bring safety risks.

[0004] The tunnel environment is dim and the span is long. The surface color of the tunnel chute is dark and the boundaries are blurred. The amount of image data collected is large and there are poorly lit and blurred images. However, labeling and annotating the chute images of railway tunnels requires a long period of manual measurement, which wastes a lot of time for professional technicians, and invisibly increases the difficulty of identifying and distinguishing defective chutes. By sorting out the data, it is found that the collected chute images often observe defects in the chute and nearby important areas, and the defects are not limited to fixed positions, but may appear around the chute. Images depicting chute defects usually show defect characteristics in multiple areas. Summary of the invention

[0005] In order to solve the above problems, the purpose of the present invention is to provide a method for detecting tunnel chute defects based on FDEResUNet network.

[0006] The present invention provides a method for detecting tunnel chute defects based on the FDEResUNet network, the method comprising: Collect the tunnel chute image to be inspected; 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; The chute defect detection model is used to output a target tunnel chute image having a defect target category position corresponding to the tunnel chute image to be detected.

[0007] Optionally, the process of training the chute defect detection model includes: Collecting sample images of the tunnel chute, generating a sample data set based on the sample images, and segmenting the sample data set to obtain a training set and a test set; Constructing a chute defect detection model based on the improved FDEResUNet network, and inputting the sample images of the training set into the chute defect detection model to train the chute 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 results.

[0008] Optionally, the chute defect detection model includes: Encoder layer: The first encoder layer that the input tunnel chute image to be detected passes through uses 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 will be used as the input of the frequency domain module FDB. A part of Li will be added to the jump connection, and the other part Hi will be used as the input of the next downsampling residual block R. Finally, the output result H4 of the encoder is input to the transfer block; Relay block: The relay block contains 1024 feature channels and is used to connect the encoder layer and the decoder layer; Decoder layer: The data of the transfer block is input into the first upsampling convolution block C of the decoder layer. The decoder stage contains four upsampling convolution blocks C. The jump connection takes the output Li of each downsampling residual block R in the encoder through the frequency domain module FDB and the output Ji from the previous encoder layer as input, and the output result is Ji+1; Double-layer enhancement block module DLE: The output J4 of the last upsampling module of the decoder layer is used as the input of the double-layer enhancement module DLE. After two double-layer enhancement modules DLE output the final enhancement result K as the final output of the model.

[0009] Optionally, the two-layer enhancement module has two parallel layers, each layer includes a two-dimensional convolution Conv, a Dropout of 0.1, a batch normalization BN, and a ReLu activation function.

[0010] Optionally, the downsampling residual block R includes a two-dimensional convolution Conv, a batch normalization BN, a Dropout of 0.3, a ReLu activation function, and the number of feature channels is doubled.

[0011] The present invention also provides a device for detecting tunnel chute defects based on the FDEResUNet network, the device comprising 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 any of the above-mentioned methods according to the instructions in the program code.

[0012] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store program codes, and the program codes are used to execute any of the above-mentioned methods.

[0013] The present invention also provides a computing device, which includes the device for detecting tunnel chute defects based on the FDEResUNet network.

[0014] The method for detecting tunnel chute defects based on the FDEResUNet network of the present invention realizes image detection by adopting the FDEResUNet network. Firstly, an improved wavelet packet transform module and a double-layer enhancement block are used to optimize the single jump link function in the original UNet structure, so as to enhance the frequency information existing in the chute image and identify the boundary of the defective target; secondly, the collected tunnel chute images are processed based on the improved FDEResUNet network. The model inputs the original chute image, passes through the encoder layer, the decoder layer and the double-layer enhancement block, and finally obtains the chute defect target segmentation result.

[0015] Based on the following detailed description of specific embodiments of the present invention in conjunction with the accompanying drawings, those skilled in the art will become more aware of the above and other objects, advantages and features of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings: Figure 1 A method for detecting tunnel chute defects based on the FDEResUNet network according to an embodiment of the present invention; Figure 2 is a schematic diagram of the process of training a chute defect detection model according to an embodiment of the present invention; Figure 3 It is a structural schematic diagram of a chute defect detection model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The embodiments of the present invention are described below in conjunction with the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the present invention and are not limiting.

[0018] An embodiment of the present invention provides a method for detecting tunnel chute defects based on a FDEResUNet network, the method comprising: S1, collecting the image of the tunnel chute to be inspected; S2, 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; S3, using the chute defect detection model to output a target tunnel chute image having a defect target category position corresponding to the tunnel chute image to be detected.

[0019] In practical applications, we can first collect chute images from the tunnel, and try to avoid dark environments during the collection process; then, the collected chute images are uniformly input into the FDEResUNet model in a 256×256 format for defect detection; finally, the chute images with segmented defective parts are output in a 256×256 format.

[0020] In this example, the chute images are classified into two categories: defective and normal. The types of defects include channel deviation, channel deformation, channel foreign matter, concrete cover, concrete missing, cement unevenness, zinc coating damage, and buried depth.

[0021] The advantages and beneficial effects of the embodiments of the present invention compared to traditional solutions are as follows: through automated image processing technology, the need for manual inspections is reduced, thereby significantly improving detection efficiency; by using a deep learning model, automated image processing technology is used to reduce the need for manual inspections, thereby significantly improving detection efficiency; the automated detection system is not limited by human vision and can cover a wider area, reducing detection 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 cost of railway tunnels in the long run.

[0022] In some optional embodiments, combined with Figure 2 As shown in the figure, the training process of the chute defect detection model can be as follows: S11 collects sample images of the tunnel chute, generates a sample data set based on the sample images, and segments the sample data set to obtain a training set and a test set; S12, constructing a chute defect detection model based on the improved FDEResUNet network, and inputting the sample images of the training set into the chute defect detection model to train the chute defect detection model; S13, inputting the tunnel chute image in the test set into a trained chute defect detection model to output a chute defect target segmentation result.

[0023] like Figure 3As shown, the chute defect detection model structure of this embodiment may include the following modules: 1. Encoder layer.

[0024] The input tunnel chute image to be detected first passes through the encoder layer using four ResNet Block 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 will be used as the input of the frequency domain block (FDB). A part of Li will be added to the jump connection, and the other part Hi will be used as the input of the next downsampling residual block R; finally, the encoder output result H4 is input to the transfer block.

[0025] The down-sampling residual block R contains two-dimensional convolution Conv, batch normalization BN, Dropout 0.3, ReLu activation function, and the number of feature channels is doubled.

[0026] 2. Frequency domain module.

[0027] The frequency domain module of this embodiment uses wavelet packet transform. Wavelet packet transform (WPT) is a signal analysis method based on wavelet transform. It can provide more flexible frequency analysis capabilities than traditional wavelet transform and decompose images at any level. The advantage of wavelet packet transform is that it does not limit the number of decomposition levels and bands, and it allows all bands to be further decomposed in each decomposition step. It is widely used in the field of image processing. Wavelet packet transform can be used to extract image features, such as edges, textures, etc. Detailed extraction of these features is crucial for this task.

[0028] In order to further improve the training effect, retain the basic information of the image and reduce the feature dimension, two-dimensional wavelet packet decomposition is used. Since tunnel chutes are mostly regular horizontal and vertical strips, the focus is on the horizontal and vertical dimensions, and wavelet packet decomposition is emphasized.

[0029] In Python, wavelet packet transform can be implemented through a variety of 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 multiple wavelet transforms, such as wavelet packet transform (WPT). The WPT function is used to perform wavelet packet transform on the output image features Ii of each downsampled residual block. The output of WPT is a waveletPacket2D object that stores all possible wavelet packet decomposition subband information in the image.

[0030] Each decomposition step will produce more sub-bands. In the first-level decomposition, the original image is decomposed into four sub-bands: LL (low frequency), LH (horizontal high frequency), HL (vertical high frequency), HH (diagonal high frequency), and the HH sub-band that causes noise interference is actively deleted; in the second-level decomposition, the LL (low frequency) sub-band is no longer decomposed, and LL is used 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 (horizontal sub-high frequency), LH_HL (vertical sub-high frequency). The HL (vertical high frequency) sub-band is further decomposed to obtain HL_LL (lower frequency vertical high frequency), HL_LH (horizontal sub-high frequency), HL_HL (vertical sub-high frequency).

[0031] Finally, 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; unify the three subbands LL, LH_LL, and HL_LL as Li, and restore the output feature map Ji+1 after element-by-element addition of the last upsampled feature map Ji with the decoder through skip connection. (In image processing, approximate coefficients are usually used to represent the rough outline of an image. In the PyWavelets library, approximate coefficients can be obtained through WPT. Approximate coefficients can be obtained by accessing specific subband paths (e.g., 'a'*level, where level refers to level)).

[0032] 3. Transfer block.

[0033] The relay block contains 1024 feature channels and is used to connect the encoder layer and the decoder layer.

[0034] As the key part that connects the encoder layer and the decoder layer together, the relay block is an essential component of the UNet network structure and can enhance the ability to extract features. The relay block involved in the model contains 1024 feature channels, including 2D convolution (Conv), batch normalization (BN), and ReLu activation function.

[0035] 4. Decoder layer.

[0036] The data of the transfer block is input into the first upsampling convolution block C of the decoder layer. The decoder stage contains four upsampling convolution blocks C. The jump connection takes the output Li of each downsampling residual block R in the encoder through the frequency domain module FDB and the output Ji from the previous encoder layer as input. The output result is Ji+1, The decoder integrates the feature information of the encoder through multiple upsampling blocks to gradually restore the size of the feature map and finally outputs the required output feature map. The upsampling stage consists of three convolution blocks with an upsampling step of 2. Each convolution block (CB) contains a 2-dimensional 3×3 kernel convolution (Conv), batch normalization (BN), and dropout of 0.2. The number of feature channels is halved (512, 256, 128, 64). The skip connection transmits the spatial information of each downsampled layer in the encoder to the upsampling decoder layer of each layer after passing through the wavelet packet transform module and the lightweight cross attention module.

[0037] 5. Dual layer enhancement block (DLE) The output J4 of the last upsampling module of the decoder layer is used as the input of the double-layer enhancement module DLE. After two double-layer enhancement modules DLE output the final enhancement result K as the final output of the model.

[0038] Due to improper operation during the installation of the tunnel chute, there will be small defects near the chute. In order to better identify these defects in the low-resolution chute images collected in the dim tunnel, a double-layer enhancement module DLE is added after the decoder layer. The double-layer enhancement module has two parallel layers, each of which contains 2D convolution (Conv), Droupout 0.1, batch normalization (BN), and ReLu activation function. The addition of the double-layer enhancement module helps the model expand the pixel range and consider more contextual information when analyzing images.

[0039] In order to achieve the goal of accurately segmenting defects in tunnel chute images, it is necessary to build a model that can recognize local features. The UNet structure has flexibility and powerful segmentation capabilities, and variants of UNet have also been widely 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 in the UNet back propagation, which helps to train the model more efficiently.

[0040] However, due to the traditional connection, the depth is difficult to estimate. If the depth is too deep, the gradient will be too small and the training process will be slowed down. In order to overcome this problem, the embodiment of the present invention adds wavelet packet transform (WPT) to the traditional UNet model, that is, the frequency domain module is used to replace the original direct connection encoder and decoder to strengthen the frequency information in the slide image.

[0041] The encoder and decoder of the ordinary UNet structure are composed of convolution blocks. We optimize the downsampling stage of the encoder to a residual block (ResNet Block, RB), which contains 2D convolution (Conv), batch normalization (BN), dropout of 0.3, and ReLu activation function. The number of feature channels is doubled (64, 128, 256, 512). Each downsampling block designed in this way will make up for the defect of gradient vanishing in the original convolution block as the training depth deepens, and also accelerate the convergence of the model. This is crucial to improve the accuracy of target detection.

[0042] A Dropout layer is added to each residual block, and the downsampling stage is set to 0.3 to prevent the same feature from being trained over and over again. The upsampling stage is set to 0.2, and Dropout is added continuously to suppress overfitting of model training. Adding the Dropout layer increases the randomness of the model training data. The ResNet block is combined with Dropout, which greatly helps the entire model training to extract the boundary of the defective part of the chute.

[0043] The tunnel chute defect detection method based on the improved FDEResUNet network provided by the embodiment of the present invention generates feature maps Ii (i=1, 2, 3, 4) in the chute image to be detected through downsampling processing at the encoder layer. After each upsampling block, the decoder restores the output feature map Ji (i=1, 2, 3, 4) after adding the upsampled feature map to the feature map from the WPT element by element. The double-layer enhancement block takes the feature map J4 output by the decoder layer as input and finally outputs the enhanced feature map K. Finally, the model outputs the result to locate the tunnel chute image that marks the location of the defect target category.

[0044] The embodiment of the present invention also provides a device for detecting tunnel chute defects based on the FDEResUNet network, the device comprising 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 for detecting tunnel chute defects based on the FDEResUNet network of the above embodiment according to the instructions in the program code. This embodiment names the chute defect detection algorithm "FDEResUNet", which involves the Res structure, F is the frequency domain block, DE is the double-layer enhancement block, these two modules are introduced in the technical disclosure book; UNet is a classic image segmentation network.

[0045] An embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store program codes, and the program codes are used to execute the method for detecting tunnel chute defects based on the FDEResUNet network of the above embodiment.

[0046] An embodiment of the present invention further provides a computing device, which includes the apparatus for detecting tunnel chute defects based on the FDEResUNet network of the above embodiment.

[0047] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A method for detecting tunnel chute defects based on FDEResUNet network, characterized in that: The method comprises: Collect the tunnel chute image to be inspected; The tunnel chute image to be detected is input into a pre-trained chute defect detection model; the chute defect detection model is constructed based on an improved FDEResUNet network; the chute defect detection model includes: an encoder layer, a transfer block, a decoder layer and a double-layer enhancement module; The chute defect detection model is used to output a target tunnel chute image having a defect target category position corresponding to the tunnel chute image to be detected.

2. The method according to claim 1, characterized in that The process of training the chute defect detection model includes: Collecting sample images of the tunnel chute, generating a sample data set based on the sample images, and segmenting the sample data set to obtain a training set and a test set; Constructing a chute defect detection model based on the improved FDEResUNet network, and inputting the sample images of the training set into the chute defect detection model to train the chute 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 results.

3. The method according to claim 1, characterized in that: Encoder layer: The first encoder layer that the input tunnel chute image to be detected passes through uses four ResNet Block 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 will be used as the input of the frequency domain module FDB. A part of Li will be added to the jump connection, and the other part Hi will be used as the input of the next downsampling residual block R. Finally, the output result H4 of the encoder is input to the transfer block; Transfer block: used to connect the encoder layer and the decoder layer; Decoder layer: The data of the transfer block is input into the first upsampling convolution block C of the decoder layer. The decoder stage contains four upsampling convolution blocks C. The jump connection takes the output Li of each downsampling residual block R in the encoder through the frequency domain module FDB and the output Ji from the previous encoder layer as input, and the output result is Ji+1; Double-layer enhancement block module DLE: The output J4 of the last upsampling module of the decoder layer is used as the input of the double-layer enhancement module DLE. After two double-layer enhancement modules DLE output the final enhancement result K as the final output of the model.

4. The method according to claim 3, characterized in that The double-layer enhancement module has two parallel layers, each of which includes a two-dimensional convolution Conv, a Dropout of 0.1, a batch normalization BN, and a ReLu activation function.

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

6. The method according to claim 3, characterized in that The down-sampling residual block R contains two-dimensional convolution Conv, batch normalization BN, Dropout 0.3, ReLu activation function, and the number of feature channels is doubled.

7. A device for detecting tunnel chute defects based on FDEResUNet network, characterized in that: The device comprises 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 to 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 codes, and the program codes are used to execute the method according to any one of claims 1 to 6.

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

Images