Image expansion method and system for underwater revetment and anchorage of navigation channel
By using an improved CBAM-Cycle-WGAN-GP style transfer model and a manually designed geometric transformation method, the problem of poor underwater image data quality in inland waterways was solved, and efficient data augmentation of small targets such as bank protection and anchorages was achieved, thereby improving image accuracy and imaging effect.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-26
- Publication Date
- 2026-03-20
AI Technical Summary
Existing technologies for underwater structure detection in inland waterways suffer from inconsistent underwater image data quality, low data volume, insufficient generalization ability and accuracy of neural network models, and difficulty in effectively expanding underwater image data using traditional methods, especially for small targets such as revetments and anchorages.
An improved CBAM-Cycle-WGAN-GP style transfer model combined with a manually designed geometric transformation method is used to generate a high-quality underwater revetment and anchorage image dataset through acquisition, preprocessing, noise removal, optical image screening, and style transfer, thereby reducing the influence of sonar attitude.
It improves the accuracy and imaging effect of underwater image data, especially the data augmentation effect for small targets such as revetments and anchorages, thereby enhancing the accuracy and robustness of the model.
Smart Images
Figure CN121708455A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of digital waterway and smart services technology, specifically to an image augmentation method and system for underwater revetments and anchorages in waterways. Background Technology
[0002] The detection of underwater structures in inland waterways has always been a key technical challenge in waterway management and maintenance. Early underwater structure detection relied primarily on manual surveying. Later, new underwater detection technologies, mainly sonar, were developed. This technology can effectively collect image data containing the morphological characteristics of underwater structures (such as navigational obstructions, revetments, and dams), providing decision support for the scientific maintenance of waterways. However, due to the limitations of sonar equipment's technical parameters and the constraints and uncertainties in engineering applications, the quality of underwater image data collected in field experiments varies greatly, and the data volume is low. Furthermore, the processing and analysis of field image data is a crucial aspect of data application. These issues, to some extent, restrict the dynamic perception and scientific maintenance of the service performance of inland waterway infrastructure. Existing research mostly uses methods such as neural network models to augment underwater image data, aiming to enrich the diversity of underwater datasets for inland waterways. However, the generalization ability, accuracy, and robustness of the models are limited by the sample size and quality of the input dataset, making it difficult to guarantee the effectiveness of data augmentation. Meanwhile, research on underwater targets needs to be deepened, with most studies focusing on large targets such as aircraft and ships; at the same time, sonar attitude affects the quality of underwater image data, such as differences in angle and orientation, different imaging on the port and starboard sides, and incomplete imaging. For the above phenomena, traditional neural network models are difficult to achieve better data augmentation effects than manually designed models. Summary of the Invention
[0003] The technical problem to be solved by the present invention is to provide an image augmentation method and system for underwater revetments and anchorages in waterways, which addresses the shortcomings of the prior art. The method realizes underwater image data augmentation by improving the neural network model, and at the same time reduces the imaging influence of sonar attitude by using artificial design methods, thereby improving accuracy and imaging effect.
[0004] To achieve the above objectives, according to one aspect of the present invention, an image augmentation method for underwater revetments and anchorages in waterways is provided, comprising: Collect raw sonar image data of underwater revetments and anchorages in the waterway to construct an initial sonar image dataset; The initial sonar image dataset is subjected to noise analysis and preprocessing to obtain a preprocessed image dataset; Optical images of anchors were collected from the internet, filtered, and preprocessed to obtain an anchor image dataset; The classic Cycle-GAN model is improved by using the CBAM (Convolutional Block Attention Module) mechanism, and the Wasserstein distance and gradient penalty (GP) mechanism is introduced to construct the CBAM-Cycle-WGAN-GP style transfer model. The preprocessed image dataset and the anchor image dataset are input into the CBAM-Cycle-WGAN-GP style transfer model to perform style transfer and generate the first expanded image set. The first expanded image set is augmented a second time using a manually designed geometric transformation method to generate the final image dataset of the revetment and anchorage.
[0005] The above scheme also includes: selecting the Inception Score (IS) index and the Fréchet Inception Distance (FID) index to evaluate the constructed style transfer model, in order to verify the effectiveness of the model improvement.
[0006] In the above scheme, the CBAM-Cycle-WGAN-GP style transfer model includes: The CBAM mechanism is introduced into the classic Cycle-GAN model to enhance feature extraction capabilities; The Wasserstein distance is introduced to replace the original GAN's cross-entropy loss, and the gradient penalty mechanism is introduced to replace weight pruning, thereby improving model stability. Based on the introduction of Wasserstein distance, both the discriminator and generator in the CBAM-Cycle-WGAN-GP style transfer model are subject to gradient penalty mechanism, and are forced to implement Lipschitz conditions and remain stable in order to control the gradient norm to fluctuate within a reasonable range. The discriminator's ability to distinguish between real and generated samples is enhanced, enabling it to more accurately guide the generator to produce high-quality images. With feedback from the discriminator, the generator continuously optimizes its generation strategy, making the generated images closer to real anchor images in terms of detail and overall style. In the style transfer model, the gradient flow of the loss function improved by the Wasserstein distance is smoother, thus improving the quality of the generated samples.
[0007] In the above scheme, the secondary image augmentation of the first augmented dataset using the manually designed geometric transformation method includes: The images in the first augmented dataset are processed using manually designed cropping, mirroring, and rotation techniques, and all resulting images are then normalized in size.
[0008] In the above scheme, the acquisition of raw sonar image data of underwater revetments and anchorages in the waterway includes: By dynamically adjusting the acquisition path and equipment parameters based on the waterway grade, hydrological and meteorological conditions, and relying on unmanned surface vessels to carry out initial image acquisition work in the actual waterway, the original sonar image data of the underwater revetment and anchorage of the waterway were obtained.
[0009] In the above scheme, the dynamic adjustment of the data acquisition path and equipment parameters based on waterway grade, hydrological and meteorological conditions specifically includes: When the air is clear and the lighting is suitable, the optimal configuration parameters of the side-scan sonar are used for data acquisition. When there are many vessels in the waterway, the measurement range is reduced and the data acquisition path is adjusted in real time to cope with changes in ship waves and water depth in the channel. By setting acquisition time windows in accordance with seasonal changes and monitoring and correcting water temperature using acoustic profiling technology, the impact of environmental factors on image quality can be reduced.
[0010] In the above scheme, the optimal configuration parameters of the side-scan sonar include: frequency of 500~700kHz, pixel accuracy of 1cm×1cm, vertical beamwidth of 45° (3dB width), and operating speed of 3 stops (1 stop = 1.852km / h).
[0011] In the above scheme, the noise analysis and preprocessing of the initial sonar image dataset includes: grayscale processing, noise type identification, and denoising processing, specifically: The images in the initial sonar image dataset are processed to grayscale, and the grayscale histograms are analyzed to identify the noise type; An adaptive filtering algorithm is used for noise reduction, which includes mean filtering and median filtering. Peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are used as evaluation metrics for denoising performance.
[0012] In the above scheme, the adaptive filtering algorithm uses mean filtering combined with a 3×3 filtering template.
[0013] According to another aspect of the present invention, an image augmentation system for underwater revetments and anchorages in waterways is provided, comprising: The acquisition module is used to acquire raw sonar image data of underwater revetments and anchorages in the waterway and to construct an initial sonar image dataset. The preprocessing module is used to perform noise analysis and preprocessing on the initial sonar image dataset to obtain a preprocessed image dataset; The collection module is used to collect optical images of anchors from the network and perform filtering and preprocessing to obtain an anchor image dataset; An improved construction and expansion module is used to improve the classic Cycle-GAN model using the CBAM mechanism, and introduces Wasserstein distance and gradient penalty mechanism to construct the CBAM-Cycle-WGAN-GP style transfer model; the preprocessed image dataset and the anchor image dataset are input into the CBAM-Cycle-WGAN-GP style transfer model for style transfer, generating the first expanded image set; The second expansion module is used to perform secondary image expansion on the first expanded image set based on a manually designed geometric transformation method, to generate the final revetment and anchorage image dataset.
[0014] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: (1) This invention provides an image augmentation method for underwater revetments and anchorages in waterways. This method achieves underwater image data augmentation by improving the neural network model, and at the same time reduces the imaging influence of sonar attitude by using artificial design methods, thereby improving accuracy and imaging effect.
[0015] (2) The present invention provides an image augmentation method for underwater revetments and anchorages in waterways. This method differs from traditional techniques that generalize underwater targets or limit them to large targets such as aircraft and ships. Instead, it targets underwater images of inland waterway infrastructure such as underwater revetments and anchorages, which have indistinct features, to achieve data augmentation. At the same time, this method combines the structural characteristics of underwater revetments and anchorages in inland waterways with underwater image features, and comprehensively considers the characteristics of possible special working conditions to improve the neural network model. It also reduces the imaging influence of sonar attitude by using artificial design methods to maximize the accuracy and imaging effect of the style transfer model. Attached Figure Description
[0016] 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: Figure 1 This is a flowchart of an image augmentation method for underwater revetments and anchorages in waterways, according to Embodiment 1 of the present invention.
[0017] Figure 2 This is a schematic diagram of the standard image acquisition route in Embodiment 1 of the present invention. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0019] It should be understood that the sequence number of each step in the embodiment does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.
[0020] Example 1 This application provides an image augmentation method for underwater revetments and anchorages in waterways. Please refer to [link to relevant documentation]. Figure 1 ,include: S1: Collect raw sonar image data of underwater revetments and anchorages in the waterway to construct an initial sonar image dataset.
[0021] Specifically, in this embodiment, the raw sonar image data of the underwater revetment and anchorage of the waterway includes: The acquisition path and equipment parameters were dynamically adjusted based on the waterway grade, hydrological and meteorological conditions. Initial image acquisition was carried out on the waterway using unmanned surface vessels. The acquired images were classified according to the underwater target categories, which included underwater revetments and anchorages. Raw sonar image data of underwater revetments and anchorages were obtained.
[0022] This embodiment involves data collection path planning and equipment parameter adjustments based on different waterway grades, hydrological conditions, and weather conditions to improve the accuracy and efficiency of data collection. The specific design is as follows: (1) When the air is clear and the lighting is suitable, the data acquisition equipment parameters are configured optimally, as shown in Table 1. For different waterway grades, the main influences on the acquisition experiment are the ship waves of navigating vessels and the water depth of the waterway. When the navigating vessels are too dense, a strategy of reducing the range is adopted. At the same time, the data acquisition path changes in real time according to the changes in the water depth below the waterway through which the experimental vessel is equipped with a single beam for real-time monitoring. A schematic diagram of the standard image acquisition route is shown in Table 1. Figure 2 As shown.
[0023] Table 1 Optimal Configuration Parameters for Side-Scan Sonar
[0024] (2) Secondly, considering the seasonal changes in the experimental area, especially the characteristics of frequent fog in spring, strong sunlight in summer, large water temperature changes in autumn, and low water temperature in winter, suitable acquisition time windows for each season were set. Acoustic profiling was used in the experimental waterway to extract parameters of the underwater water temperature, determine the impact of water temperature changes on the experimental data, and provide corrections to reduce the impact of environmental factors on image quality. In response to changes in underwater lighting conditions, such as light attenuation due to water depth and interference from water surface reflection, the exposure time, gain, and other parameters of the acquisition equipment were dynamically adjusted to ensure that the acquired images were clear, detailed, and of high quality.
[0025] S2 performs noise analysis and preprocessing on the initial sonar image dataset to obtain the preprocessed image dataset.
[0026] Given that raw image data is susceptible to interference from acquisition operations and equipment factors, leading to noise issues and blurring of object features in the image, which negatively impacts the effective use of subsequent models, this step requires noise analysis of the initial sonar image dataset. In this embodiment, noise analysis and preprocessing of the initial sonar image dataset includes: grayscale processing, noise type identification, and denoising. The denoising effect is then illustrated using quantization standards, as detailed below: Based on the sonar imaging principle and the grayscale histogram of the image, it was determined that the distribution characteristics of the noise were consistent with those of salt-and-pepper noise. Therefore, an adaptive filtering algorithm (including mean filtering and median filtering) was adopted to optimize the salt-and-pepper noise and low-frequency interference in underwater structure images. The method selection was determined by comparing the effects and using objective evaluation indicators.
[0027] The formula for mean filtering is as follows:
[0028] in, Therefore The pixel values of the center's neighborhood; These are the weighting coefficients for each neighboring pixel, where is the template size used. , This represents the boundary range of pixel values for which the mean value is to be used.
[0029] The formula for median filtering is as follows:
[0030] in, f ( x , y )and g ( x , y ) represent the pixel values in the original image and the processed image, respectively. S For windows.
[0031] The objective evaluation metrics selected are Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM), and their calculation formulas are as follows:
[0032] in, f i,j Represents the pixel value information of the original real image; M , N These represent the length and width values of the image, respectively. f’ i,j This indicates the processed image pixel value information.
[0033]
[0034] in, m This indicates calculating the mean of the corresponding image; s This indicates the calculation of standard deviation; s xy express x , y The covariance between them. c 1. c Both 2 are constant coefficients used to prevent the denominator from being zero, and are calculated as follows:
[0035]
[0036] in, L This refers to the dynamic range of grayscale, which is related to the type of image data. Here, we take 255. K 1. Take a constant value of 0.01. K 2 is set in the code, and is taken as 0.03.
[0037] Specifically, in this embodiment, mean filtering is used in conjunction with its 3×3 filter template to obtain data with less noise and higher clarity.
[0038] S3: Collect optical images of anchors from the network, filter and preprocess them to obtain an anchor image dataset.
[0039] Specifically, in this embodiment, since underwater targets such as anchors appear infrequently in experiments, the features of anchors in optical images collected from the internet are input into the model for feature learning. This enables the model to recognize and understand the image features of anchors under different environmental and lighting conditions. At the same time, to ensure the quality and applicability of the input images, the optical images collected from the internet are screened and preprocessed. Specifically, images that are irrelevant to the features of anchors or of low quality are removed, resulting in an anchor image dataset for the model to use, thereby improving the accuracy and efficiency of the model's learning.
[0040] S4 utilizes the CBAM mechanism to improve the classic Cycle-GAN model and introduces Wasserstein distance and gradient penalty mechanisms to construct the CBAM-Cycle-WGAN-GP style transfer model. The preprocessed image dataset and anchor image dataset are input into the CBAM-Cycle-WGAN-GP style transfer model for style transfer, generating the first expanded image set.
[0041] In this embodiment, due to limitations in the field experimental sample collection and the limitations of traditional augmentation algorithm models, the classic Cycle-GAN model is improved using the CBAM (Convolutional Block Attention Module) mechanism to increase the model's ability to extract image features. Since the target occupies a small proportion of the image, and the addition of CBAM enhances the discriminator's feature extraction ability, it can lead to more severe gradient vanishing in the generator. Therefore, this embodiment also introduces the Wasserstein distance and gradient penalty (GP) mechanism to construct a CBAM-Cycle-WGAN-GP style transfer model, and uses the existing anchor image dataset as the target for style transfer generation. While enhancing the quality and resolution of the generated images, it ensures that the model can stably reach convergence. Style transfer is then performed on the preprocessed image dataset to obtain the first augmented image set.
[0042] The style transfer model involved in the image data augmentation part of this embodiment includes the following specific functions: Style transfer model of the classic generative adversarial network Cycle-GAN; CBAM mechanism network model – improves the model's feature extraction capability; Introducing Wasserstein's GAN model—replacing the original GAN's cross-entropy loss to improve model stability; Gradient penalty mechanism model – an alternative to weight pruning, improving model stability; Based on the improved CBAM-Cycle-WGAN-GP style transfer model using CBAM combined with GP; Style transfer models involve discriminators and generators, and the specific formula is as follows:
[0043]
[0044] in: The input dataset (i.e., the preprocessed image dataset); To generate a dataset; Generator in a positive loop Combating losses Generator for reverse loop The adversarial loss is used to reconstruct and enhance the input set; and To determine whether a sample belongs to and Two discriminators; and To implement respectively X Style transformed Style and Implementation Style transformed Two style generators; For distributions from the real domain samples y Expectations For distributions from the real domain samples x The expectation.
[0045] The calculation formula for the CBAM mechanism is as follows:
[0046] in: M CBAM For channel attention; MLP Represented as a multilayer perceptron; AP This is represented as average pooling; MP This is represented as max pooling; F This is represented as the input feature map. d This is represented as the weight of the mapping.
[0047] The formula for the gradient penalty term is:
[0048]
[0049] in:
[0050]
[0051] The weight coefficient for gradient penalty is set to 10 in this embodiment.
[0052] Based on the above, the total discriminator loss, including gradient penalty, is obtained: Discriminator D Y Total loss:
[0053] Discriminator D X Total loss:
[0054] The total loss of the generator is as follows:
[0055] The cycle consistency loss is:
[0056]
[0057] S5 uses a manually designed geometric transformation method to perform secondary image augmentation on the first augmented image set, generating the final revetment and anchorage image dataset.
[0058] Understandably, given that the first expansion involves style transfer, it is difficult to effectively expand image data with specific imaging characteristics (such as differences in sonar angle and azimuth, port and starboard imaging characteristics, and incomplete imaging features). Therefore, this embodiment employs three manually designed geometric transformations—rotation, cropping, and mirroring—to perform a second image dataset expansion on the first expanded image set, generating the final revetment and anchorage image dataset. These three processing methods respectively simulate changes in sonar towing angle and azimuth, local burial phenomena caused by sonar imaging obstruction, and different imaging modes on the port and starboard sides of the sonar. All obtained images are then normalized in size to facilitate the application of a unified dataset format in subsequent steps.
[0059] In summary, this invention provides an image augmentation method for underwater revetments and anchorages in waterways. This method augments underwater image data by improving a neural network model, and supplements image features that affect sonar attitude with artificial design methods, thereby improving accuracy and imaging effect.
[0060] Example 2 This application provides an image augmentation method for underwater revetments and anchorages in waterways. The method provided in this embodiment is basically the same as the method provided in Embodiment 1, except that the method in this embodiment further includes: The Inception Score (IS) and Fréchet Inception Distance (FID) indices were selected to evaluate the constructed style transfer model and to verify the effectiveness of the model improvement.
[0061] In this embodiment, to objectively compare the effects of the model before and after improvement, the Inception Score (IS) and Fréchet Inception Distance (FID) indices are selected to determine the improvement effect and rationality of the model improvement.
[0062] The calculations for the two indices are as follows:
[0063]
[0064] Among them: image x The sampling operation is based on the distribution of the generated sonar image. P g In the process, KL For divergence, p ( g | x () represents the classification vector obtained after inputting the generated sonar image into the Inception Net-V3 model. It is the edge distribution of the generated image; and These represent the optical image and the generated sonar image, respectively. m x and m g The mean of the image features. l x and l g Let covariance be the image feature matrix. Tr The trace of the matrix; Indicates the distribution from generated sonar images P g samples x The expectation.
[0065] Another embodiment of this application provides an image augmentation system for underwater revetments and anchorages in waterways, including: The acquisition module is used to acquire raw sonar image data of underwater revetments and anchorages in the waterway and to construct an initial sonar image dataset. The preprocessing module is used to perform noise analysis and preprocessing on the initial sonar image dataset to obtain the preprocessed image dataset; The collection module is used to collect optical images of anchors from the network and perform filtering and preprocessing to obtain an anchor image dataset; An improved construction and expansion module is used to improve the classic Cycle-GAN model using the CBAM mechanism, and introduces Wasserstein distance and gradient penalty mechanism to construct the CBAM-Cycle-WGAN-GP style transfer model; the preprocessed image dataset and anchor image dataset are input into the CBAM-Cycle-WGAN-GP style transfer model for style transfer, generating the first expanded image set; The second expansion module is used to perform secondary image expansion on the first expanded image set based on a manually designed geometric transformation method, generating the final revetment and anchorage image dataset.
[0066] It should be noted that, depending on the implementation needs, the various steps described in this application can be broken down into more steps, or two or more steps or parts of the steps can be combined into new steps to achieve the purpose of this invention.
[0067] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. An image augmentation method for underwater revetments and anchorages in waterways, characterized in that, include: Collect raw sonar image data of underwater revetments and anchorages in the waterway to construct an initial sonar image dataset; The initial sonar image dataset is subjected to noise analysis and preprocessing to obtain a preprocessed image dataset; Optical images of anchors were collected from the internet, filtered, and preprocessed to obtain an anchor image dataset; The classic Cycle-GAN model is improved using the CBAM mechanism, and the Wasserstein distance and gradient penalty mechanism are introduced to construct the CBAM-Cycle-WGAN-GP style transfer model. The preprocessed image dataset and the anchor image dataset are input into the CBAM-Cycle-WGAN-GP style transfer model for style transfer to generate the first expanded image set. The first expanded image set is augmented a second time using a manually designed geometric transformation method to generate the final image dataset of the revetment and anchorage.
2. The image augmentation method for underwater revetments and anchorages in waterways according to claim 1, characterized in that, Also includes: The Inception Score and Fréchet Inception Distance index were selected to evaluate the constructed style transfer model and to verify the effectiveness of the model improvement.
3. The image augmentation method for underwater revetments and anchorages in waterways according to claim 1, characterized in that, The CBAM-Cycle-WGAN-GP style transfer model includes: The CBAM mechanism is introduced into the classic Cycle-GAN model to enhance feature extraction capabilities; The Wasserstein distance is introduced to replace the original GAN's cross-entropy loss, and the gradient penalty mechanism is introduced to replace weight pruning, thereby improving model stability. Based on the introduction of Wasserstein distance, both the discriminator and generator in the CBAM-Cycle-WGAN-GP style transfer model are subject to gradient penalty mechanism, and are forced to implement Lipschitz conditions and remain stable in order to control the gradient norm to fluctuate within a reasonable range.
4. The image augmentation method for underwater revetments and anchorages in waterways according to claim 1, characterized in that, The manually designed geometric transformation method for secondary image augmentation of the first augmented dataset includes: The images in the first augmented dataset are processed using manually designed cropping, mirroring, and rotation techniques, and all resulting images are then normalized in size.
5. The image augmentation method for underwater revetments and anchorages in waterways according to claim 1, characterized in that, The raw sonar image data of the underwater revetment and anchorage of the waterway collected includes: By dynamically adjusting the acquisition path and equipment parameters based on the waterway grade, hydrological and meteorological conditions, and relying on unmanned surface vessels to carry out initial image acquisition work in the actual waterway, the original sonar image data of the underwater revetment and anchorage of the waterway were obtained.
6. The image augmentation method for underwater revetments and anchorages in waterways according to claim 5, characterized in that, The dynamic adjustment of the data acquisition path and equipment parameters based on waterway grade, hydrological and meteorological conditions specifically includes: When the air is clear and the lighting is suitable, the optimal configuration parameters of the side-scan sonar are used for data acquisition. When there are many vessels in the waterway, the measurement range is reduced and the data acquisition path is adjusted in real time to cope with changes in ship waves and water depth in the channel. By setting acquisition time windows in accordance with seasonal changes and monitoring and correcting water temperature using acoustic profiling technology, the impact of environmental factors on image quality can be reduced.
7. The image augmentation method for underwater revetments and anchorages in waterways according to claim 6, characterized in that, The optimal configuration parameters of the side-scan sonar include: frequency of 500~700kHz, pixel accuracy of 1cm×1cm, vertical beamwidth of 45°, and operating speed of 3 segments.
8. The image augmentation method for underwater revetments and anchorages in waterways according to claim 1, characterized in that, The noise analysis and preprocessing of the initial sonar image dataset includes: grayscale processing, noise type identification, and denoising processing, specifically: The images in the initial sonar image dataset are processed to grayscale, and the grayscale histograms are analyzed to identify the noise type; An adaptive filtering algorithm is used for noise reduction, which includes mean filtering and median filtering. Peak signal-to-noise ratio and structural similarity are used as evaluation metrics for denoising effectiveness.
9. The image augmentation method for underwater revetments and anchorages in waterways according to claim 8, characterized in that, The adaptive filtering algorithm uses mean filtering combined with a 3×3 filter template.
10. An image augmentation system for underwater revetments and anchorages in waterways, characterized in that, include: The acquisition module is used to acquire raw sonar image data of underwater revetments and anchorages in the waterway and to construct an initial sonar image dataset. The preprocessing module is used to perform noise analysis and preprocessing on the initial sonar image dataset to obtain a preprocessed image dataset; The collection module is used to collect optical images of anchors from the network and perform filtering and preprocessing to obtain an anchor image dataset; An improved construction and expansion module is used to improve the classic Cycle-GAN model using the CBAM mechanism, and introduces Wasserstein distance and gradient penalty mechanism to construct the CBAM-Cycle-WGAN-GP style transfer model; the preprocessed image dataset and the anchor image dataset are input into the CBAM-Cycle-WGAN-GP style transfer model for style transfer, generating the first expanded image set; The second expansion module is used to perform secondary image expansion on the first expanded image set based on a manually designed geometric transformation method, to generate the final revetment and anchorage image dataset.