An image denoising processing method for sewer defect detection

Abstract

The present application relates to a kind of image denoising processing methods for sewer defect detection, comprising: obtaining original pipe image data set for training, and the original pipe image data set obtained is data enhanced to obtain enhanced pipe image data set;Enhanced pipe image in enhanced pipe image data set is as clean image, randomly add noise in clean image to obtain the noise image corresponding to clean image;Noise image is input image denoising model and is processed to obtain its corresponding denoising image, and loss function is constructed according to the mean square error between the denoising image of noise image and its clean image, and the image denoising model is trained by Adam optimizer;The pipe image to be processed is input into trained image denoising model, and the pipe image after denoising processing is obtained.The present application can realize the efficient filtration of complex pipe noise, while retaining image key details, improve the accuracy of subsequent defect detection.

Description

Technical Field

[0001] This invention belongs to the field of pipeline defect detection technology, and in particular relates to an image denoising processing method for detecting defects in drainage pipelines. Background Technology

[0002] With the acceleration of urbanization, the inspection and maintenance of drainage pipe networks have become increasingly important. Drainage pipe networks not only bear the heavy responsibility of urban rainwater drainage and sewage transportation, but also involve the sustainable development of the urban environment and the guarantee of residents' quality of life. However, the drainage pipe network environment is complex, presenting the following challenges: due to the humid and poorly lit internal environment of the drainage pipe network, and the potential interference from sewage and debris, defect images often suffer from blurriness and noise, affecting image readability and analytical accuracy. Therefore, the large amount of image data collected by the inspection system needs to undergo effective denoising and enhancement processing to improve image quality and facilitate subsequent analysis and diagnosis.

[0003] In the image detection of defects in drainage pipe networks, image denoising technology is of great significance. By removing noise and blur from images, the clarity and quality of the images can be significantly improved, thereby more accurately identifying problems in the pipe network, such as blockages, leaks, and corrosion. Furthermore, high-quality image data can also help realize automated fault diagnosis and early warning systems, improving the operational efficiency and management level of drainage pipe networks. In practical applications of drainage pipe network detection, traditional image denoising methods such as mean filtering and median filtering, while simple, have limited effectiveness in handling complex noise and may lose important details. While deep learning-based denoising models, such as Generative Adversarial Networks (GANs), perform well in denoising, their high computational complexity makes them difficult to meet real-time requirements. Similarly, DnCNN, although with a much lower computational complexity than GANs, suffers from less than ideal denoising performance in complex environments due to its simpler model structure. Therefore, developing a drainage pipe image denoising method with low computational complexity and good denoising performance is of great importance. Summary of the Invention

[0004] To address the problems existing in the background art, one aspect of the present invention provides an image denoising processing method for detecting defects in drainage pipes, comprising:

[0005] S1: Obtain the original pipeline image dataset for training, and perform data augmentation on the obtained original pipeline image dataset to obtain the augmented pipeline image dataset;

[0006] S2: Take the enhanced pipe images in the enhanced pipe image dataset as clean images, and randomly add noise to the clean images to obtain the noise images corresponding to the clean images;

[0007] S3: Input the noisy image into the image denoising model to obtain its corresponding denoised image. Construct a loss function based on the mean square error between the denoised image and the clean image of the noisy image, and train the image denoising model through the Adam optimizer.

[0008] S4: Input the pipeline image to be processed into the trained image denoising model to obtain the denoised pipeline image.

[0009] Another aspect of the present invention provides a drainage pipe defect image denoising processing system, characterized in that the system includes a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the image denoising processing method for drainage pipe defect detection.

[0010] A computer storage medium, characterized in that the computer storage medium stores a remote monitoring program, which, when executed by a processor, implements the image denoising processing method for detecting defects in drainage pipes.

[0011] The present invention has at least the following beneficial effects

[0012] This invention effectively addresses the limitations of traditional mean filtering and median filtering methods in handling complex noise in drainage pipes (such as blurring caused by moisture and interference from sewage debris), which are prone to losing defect details. It also overcomes the shortcomings of DnCNN models due to their simple structure and insufficient adaptability in complex environments, as well as the high computational complexity of deep learning models like GANs, which struggle to meet real-time detection requirements. This invention enriches the diversity of training data through data augmentation strategies. The image denoising model (an improved ConvNAT network fusing ConvNeXt modules (large kernel depth separable convolutions and residual connections) and CBAM modules (channel + spatial attention mechanisms) can accurately capture noise distribution and focus on key defect features. Combined with the MSE loss function and Adam optimizer, it achieves efficient training and inference, reducing computational overhead while ensuring superior denoising performance. Furthermore, it enhances the model's generalization ability across different pipe environments, providing high-quality data support for downstream automated fault diagnosis and other processes, thus contributing to the upgrading of drainage pipe network detection and management. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0014] Figure 2 This is a schematic diagram of the image denoising model of the present invention;

[0015] Figure 3 This is a schematic diagram showing the experimental simulation comparison of the present invention. Detailed Implementation

[0016] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0017] Please see Figure 1 This invention provides an image denoising processing method for detecting defects in drainage pipes, comprising:

[0018] S1: Obtain the original pipeline image dataset for training, and perform data augmentation on the obtained original pipeline image dataset to obtain the augmented pipeline image dataset;

[0019] Preferably, the data augmentation of the acquired raw pipeline image dataset includes:

[0020] S11: Normalize the acquired raw pipe image data and scale it according to various ratios to obtain scaled pipe images of different scales.

[0021] S12: Extract 40×40 pixel image blocks by sliding the scaled pipeline images at each scale in 10-pixel increments;

[0022] S13: Apply random geometric transformations to each image patch to obtain the augmented image dataset.

[0023] In this embodiment, by acquiring the original pipeline image dataset, the pixel values ​​(0-255) of all original pipeline images (640×480 pixels) are normalized to the 0-1 range, and then scaled at four ratios of 0.8, 1.0, 1.2, and 1.5 to obtain scaled images of four scales: 512×384, 640×480, 768×576, and 960×720. Subsequently, S12 is executed: 40×40 pixel image blocks are extracted from each scale image with a step size of 10 pixels, for example, a 640×480 image. 61 locations can be extracted horizontally (640-40)÷10+1=61 locations and 45 locations can be extracted vertically (480-40)÷10+1=45 locations, for a total of 2745 image patches. The same number of image patches can be obtained for images of other scales. Finally, S13 is executed: a geometric transformation is randomly applied to each image patch (such as randomly selecting 1-2 of the following: 0° / 90° / 180° / 270° rotation, horizontal flip, and vertical flip). Finally, all transformed image patches are summarized to form an augmented pipeline image dataset containing more than 100,000 image patches. Step S1 effectively addresses the potential issues of insufficient quantity, limited scenarios (e.g., covering only a few pipe diameters and defect types), and feature limitations in the original pipeline image dataset through a multi-dimensional data augmentation strategy. It not only adapts the model to pipeline defect and noise features at different scales through normalization and multi-scale scaling, but also enriches the spatial angle and positional diversity of the data through sliding blocks and random geometric transformations. This enables the augmented dataset to comprehensively cover image features in the complex environment of drainage pipelines, providing more sufficient sample support for subsequent model training, significantly reducing the risk of model overfitting, and improving the model's generalization ability and learning effect for different pipeline scenarios, different defect types, and complex noise.

[0024] S2: Take the enhanced pipe images in the enhanced pipe image dataset as clean images, and randomly add noise to the clean images to obtain the noise images corresponding to the clean images;

[0025] This embodiment employs a noise-denoising self-supervised learning paradigm. Training sample pairs are constructed by adding synthetic noise to clean images, enabling the model to learn the mapping relationship from noisy images to the original images. Specifically, in each training batch, Gaussian white noise at a specified noise level is added to the input clean image patch to generate noisy samples, ultimately forming noisy image-clean image training pairs. This self-supervised method requires no additional noisy image annotation; training can be completed using only the original clean images themselves. The noise standard deviation σ is used as a hyperparameter to control the noise intensity. This method automates the data loading and noise injection process, enabling the model to learn general image denoising capabilities. This process is expressed by the following formula:

[0026]

[0027] in, This represents the set of noisy images after adding noise. Represents the original, clean set of images. This represents the addition of Gaussian white noise.

[0028] Please see Figure 2 S3: Input the noisy image into the image denoising model to obtain its corresponding denoised image. Construct a loss function based on the mean square error between the denoised image and the clean image of the noisy image, and train the image denoising model through the Adam optimizer.

[0029] Preferably, the image denoising model employs an improved ConvNAT network model comprising: a first convolutional layer, a first GELU activation function, an expansion layer, a contraction layer, a function enhancement layer, and a third convolutional layer, cascaded sequentially; wherein the expansion layer and the function enhancement layer employ multiple sequentially interleaved ConvNeXt modules and CBAM modules, respectively; and the contraction layer comprises a second convolutional layer, a batch normalization layer, and a second GELU activation function, cascaded sequentially.

[0030] Preferably, the ConvNeXt module includes: a large kernel depth-separable convolutional layer, a layer normalization layer, a first point convolutional layer, a third GELU activation function, and a second point convolutional layer, which are cascaded in sequence; wherein the output of the ConvNeXt module is residually connected to the output of the ConvNeXt module.

[0031] Preferably, the CBAM module includes a channel attention mechanism module and a spatial attention mechanism module cascaded in sequence; the channel attention mechanism module is used to generate a corresponding channel weight parameter matrix based on the input feature map of the CBAM module, and to weight the input feature map using the channel weight parameter matrix to obtain an intermediate feature map; the spatial attention mechanism module is used to generate a corresponding spatial weight parameter matrix based on the intermediate feature map, and to weight the input intermediate feature map using the spatial weight parameter matrix to obtain the output feature map of the CBAM module.

[0032] In this embodiment, the processing of the input image denoising model for noisy images includes: taking a batch (e.g., batchsize=32) of noisy images as input, each image being 40 pixels in size (RGB three-channel); firstly, inputting them into the first convolutional layer of the improved ConvNAT network (using a 3-kernel convolutional layer, with the output channel number set to 64) to perform preliminary feature extraction, obtaining a 64-channel feature map; then, introducing nonlinear features through the first GELU activation function to enhance the model's ability to express noise features. Next, the extended layer is entered, which consists of 8 ConvNeXt modules and 8 CBAM modules sequentially interleaved; the ConvNeXt module first performs spatial feature capture on the feature map using a 7×7 large-kernel depthwise separable convolution, and after layer normalization to unify the feature distribution, it increases the channel number to 128 through a 1×1 first-point convolution, then performs a third GELU activation, and finally reduces the dimension back to 64 channels through a 1×1 second-point convolution, while simultaneously performing a residual connection with the module input to avoid gradient vanishing; the subsequent CBAM module first uses a channel attention mechanism (to apply channel attention to the input feature map)... Global average pooling and global max pooling are performed, and the results are input into a shared fully connected layer. A 64-dimensional channel weight matrix is ​​generated through Sigmoid activation (and weighted channel-wise with the input feature map) to obtain an intermediate feature map. Then, a spatial attention mechanism is used (channel-wise average pooling and max pooling are performed on the intermediate feature map, concatenated, and then weighted pixel-wise with a 40×40 spatial weight matrix generated through 3×3 convolution and Sigmoid activation) to output enhanced features. Subsequent ConvNeXt and CBAM modules repeat the above process to gradually enhance the discriminative power between noise and defect features. Then, a shrinking layer (sequentially cascaded with three heavily convolutional layers, a batch normalization layer, and a second GELU activation function) compresses the feature dimension, focusing the core feature information. Afterward, a feature enhancement layer (composed of eight ConvNeXt modules and eight CBAM modules) further optimizes the feature representation. Finally, a third convolutional layer (three intersecting convolutional kernels, outputting three channels) restores the denoised image to its original 40-dimensional output of three channels. For the 32 sets of "clean image - denoised image" in this batch, the loss was calculated according to the MSE formula:

[0033] Preferably, the mean square error between the denoised image and its clean image of the noisy image includes:

[0034]

[0035] in, The mean square error between the denoised image and the clean image of the noisy image; This represents the clean image corresponding to the noisy image. Denoising image corresponding to noisy image; and This indicates the number of pixels in the height and width directions of the noisy image; The clean image corresponding to the noisy image is in the first position. Line 1 The pixel values ​​corresponding to the column; The denoised image corresponding to the noisy image is in the th position. Line 1 The pixel value corresponding to the column.

[0036] In this embodiment, the Adam optimizer is used (with a learning rate of 0.001, β1=0.9, β2=0.999). Based on the calculated loss value, the parameters of all convolutional layers, normalization layers, etc., in the network are updated through backpropagation. The process of "input batch - feature processing - loss calculation - parameter update" is repeated. The above scheme employs an improved ConvNAT network that integrates the ConvNeXt module (large kernel depthwise separable convolution and residual connections) and the CBAM module (channel + spatial attention mechanism). This allows it to capture the global distribution patterns of complex noise in pipeline images using large kernel convolution, and to focus on key details of defects such as blockage and corrosion through the attention mechanism. This approach avoids losing important features during denoising, addressing the limitations of traditional filtering methods in handling complex noise and the insufficient adaptability of the simple DnCNN model structure. Furthermore, by constructing a loss function based on the mean squared error (MSE) between noisy and clean images, the denoising error can be accurately quantified. Combined with the Adam optimizer, this enables rapid model convergence, shortening the training cycle. Compared to the computationally complex GAN model, this model reduces computational overhead through lightweight design, balancing denoising performance and efficiency. The resulting model efficiently outputs high-quality denoised images, providing clear and accurate image data support for subsequent drainage pipe defect detection, thus improving overall detection accuracy and reliability.

[0037] S4: Input the pipeline image to be processed into the trained image denoising model to obtain the denoised pipeline image.

[0038] Experimental simulation:

[0039] Please see Figure 3 In this embodiment, 100 pipe images were selected from the Sewer-ML dataset to construct a test set. The Sewer-ML dataset contains 75,618 sewer inspection videos collected from three Danish water companies, totaling 1.3 million images. Based on the constructed test set, the SSIM, PSNR, and MACs metrics of the test set on ConvNAT network (the image denoising model proposed in this invention) and DnCNN (an existing image denoising model) were calculated. The comparison results are shown in the table below:

[0040] Table 1 Comparison of ConvNAT and DnCNN on the same test set.

[0041]

[0042] Through Table 1 and Figure 2 As can be seen, under the same conditions, ConvNAT achieves better denoising results on the test set with fewer computations. The SSIM and PSNR metrics are calculated as follows:

[0043]

[0044]

[0045] in, It is a clean image. The maximum pixel value; and This small constant was added to avoid division by zero; It is an image and images Covariance between them; and Representing images respectively and images The variance; and Representing images respectively and images The mean.

[0046] Another aspect of the present invention provides a drainage pipe defect image denoising processing system, characterized in that the system includes a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the image denoising processing method for drainage pipe defect detection.

[0047] A computer storage medium, characterized in that the computer storage medium stores a remote monitoring program, which, when executed by a processor, implements the image denoising processing method for detecting defects in drainage pipes.

[0048] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0049] In summary, this invention effectively addresses the limitations of traditional mean filtering and median filtering methods in handling complex noise in drainage pipes (such as blurring caused by moisture and interference from sewage debris), which are prone to losing defect details. It also overcomes the shortcomings of DnCNN models due to their simple structure and insufficient adaptability in complex environments, as well as the high computational complexity of deep learning models like GANs, which struggle to meet real-time detection requirements. This invention enriches the diversity of training data through data augmentation strategies. The image denoising model (an improved ConvNAT network fusing ConvNeXt modules (large kernel depth separable convolutions and residual connections) and CBAM modules (channel + spatial attention mechanisms) can accurately capture noise distribution and focus on key defect features. Combined with the MSE loss function and Adam optimizer, it achieves efficient training and inference, reducing computational overhead while ensuring superior denoising performance. Furthermore, it enhances the model's generalization ability across different pipe environments, providing high-quality data support for downstream automated fault diagnosis and other processes, thus contributing to the upgrading of drainage pipe network detection and management.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An image denoising method for detecting defects in drainage pipes, characterized in that, include: S1: Obtain the original pipeline image dataset for training, and perform data augmentation on the obtained original pipeline image dataset to obtain the augmented pipeline image dataset; S2: Take the enhanced pipe images in the enhanced pipe image dataset as clean images, and randomly add noise to the clean images to obtain the noise images corresponding to the clean images; S3: Input the noisy image into the image denoising model to obtain its corresponding denoised image. Construct a loss function based on the mean square error between the denoised image and the clean image of the noisy image, and train the image denoising model through the Adam optimizer. S4: Input the pipeline image to be processed into the trained image denoising model to obtain the denoised pipeline image; The image denoising model employs an improved ConvNAT network model, comprising: a first convolutional layer, a first GELU activation function, an expansion layer, a contraction layer, a function enhancement layer, and a third convolutional layer, cascaded sequentially; wherein the expansion layer and the function enhancement layer employ multiple sequentially interleaved ConvNeXt modules and CBAM modules, respectively; the contraction layer comprises a second convolutional layer, a batch normalization layer, and a second GELU activation function, cascaded sequentially.

2. The image denoising method for detecting defects in drainage pipes according to claim 1, characterized in that, The data augmentation of the acquired raw pipeline image dataset includes: S11: Normalize the acquired raw pipe image data and scale it according to various ratios to obtain scaled pipe images of different scales. S12: Extract 40×40 pixel image blocks by sliding the scaled pipeline images at each scale in 10-pixel increments; S13: Apply random geometric transformations to each image patch to obtain the augmented image dataset.

3. The image denoising method for detecting defects in drainage pipes according to claim 2, characterized in that, The ConvNeXt module includes: a large kernel depthwise separable convolutional layer, a layer normalization layer, a first point convolutional layer, a third GELU activation function, and a second point convolutional layer, which are cascaded in sequence; wherein, the output of the ConvNeXt module is residually connected to the output of the ConvNeXt module.

4. The image denoising method for detecting defects in drainage pipes according to claim 3, characterized in that, The CBAM module includes a channel attention mechanism module and a spatial attention mechanism module cascaded in sequence. The channel attention mechanism module is used to generate a corresponding channel weight parameter matrix based on the input feature map of the CBAM module, and to weight the input feature map using the channel weight parameter matrix to obtain an intermediate feature map. The spatial attention mechanism module is used to generate a corresponding spatial weight parameter matrix based on the intermediate feature map, and to weight the input intermediate feature map using the spatial weight parameter matrix to obtain the output feature map of the CBAM module.

5. The image denoising method for detecting defects in drainage pipes according to claim 1, characterized in that, The mean square error between the denoised image and its clean image of the noisy image includes: in, The mean square error between the denoised image and the clean image of the noisy image; This represents the clean image corresponding to the noisy image. Denoising image corresponding to noisy image; and This indicates the number of pixels in the height and width directions of the noisy image; The clean image corresponding to the noisy image is in the first position. Line number The pixel values ​​corresponding to the column; The denoised image corresponding to the noisy image is in the th position. Line number The pixel value corresponding to the column.

6. A system for denoising images of defects in drainage pipes, characterized in that, The system includes a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute an image denoising processing method for detecting defects in drainage pipes as described in any one of claims 1 to 5.

7. A computer storage medium, characterized in that, The computer storage medium stores a remote monitoring program, which, when executed by a processor, implements an image denoising processing method for detecting defects in drainage pipes as described in any one of claims 1 to 5.

Images