Acoustic-optical imaging semantic fusion detection method for underwater defects of hydraulic structure

Abstract

The invention discloses an acoustic-optical imaging semantic fusion detection method for underwater defects of a hydraulic structure, and the method comprises the steps: obtaining underwater acoustic and optical images of the hydraulic structure at the same time through underwater acoustic-optical imaging sensing for the underwater defects of the hydraulic structure; and on the basis, acoustic semantic features in the acoustic image are extracted through a semantic feature extraction module, optical image features in the optical image are processed by taking the acoustic semantic features as guidance, and a defect detection result is obtained when an underwater defect target exists in the image, so that underwater defect sound-light imaging semantic fusion detection is realized. The technology is suitable for health monitoring of various dam system hydraulic structures, and operation safety and water safety of hydraulic engineering are guaranteed.

Description

Technical Field

[0001] The present invention relates to a method for detecting underwater defects of hydraulic structures using acoustic-optical imaging semantic fusion, and in particular to a semantically guided acoustic-optical imaging fusion detection framework and method, as well as its application technology in health monitoring of various underwater engineering structures, belonging to the technical field of underwater engineering structure monitoring. Background Art

[0002] With the development of marine engineering and underwater intelligent operation and maintenance, the demand for structural health monitoring of underwater structures (such as bridge piers, pipelines, ship hulls, and offshore platforms) is becoming increasingly urgent. Their safety and integrity are directly related to the long-term serviceability and operational assurance of water conservancy projects. Traditional underwater inspection primarily relies on single-modality imaging. Especially in complex environments such as low visibility and turbid water, conventional optical imaging systems are severely affected by light attenuation and scattering, resulting in short detection ranges and numerous constraints, making them difficult to accomplish large-scale underwater defect monitoring tasks. Therefore, acoustic imaging technology, with its excellent penetration capability and stable imaging results, has been widely used in underwater target recognition and defect detection. However, single-modality acoustic imaging still faces numerous challenges, such as low resolution, blurred target boundaries, poor signal-to-noise ratio, and susceptibility to background interference, which seriously affect the robustness and accuracy of detection models. Furthermore, due to the complexity of underwater propagation media, acoustic wave propagation is often accompanied by attenuation, multipath effects, bubble interference, and seafloor sediment artifacts, further exacerbating the detection model's misjudgment rate. In recent years, computer vision and deep learning technologies have developed rapidly, and detection methods based on CNNs and Transformers have achieved breakthroughs in the field of natural images. However, sonar images differ significantly from natural images in terms of imaging mechanisms, texture representation, and visual properties. First, low contrast and blurred boundaries make it difficult for traditional image detection algorithms to accurately segment defect areas; second, strong background interference significantly increases the risk of false detection; and third, the limited information representation capability of a single modality makes it difficult to support complex defect classification and measurement tasks. Given the complementary imaging principles of acoustic and optical images—the former offers advantages such as strong penetration and stable imaging, while the latter boasts high resolution and clear textures—multimodal information fusion has gradually become a key development direction in underwater defect detection technology. Some studies have attempted to jointly model acoustic and optical images, such as using image registration or attention mechanisms to achieve preliminary fusion. However, most methods remain at a superficial level of feature splicing, struggling to deeply explore the structural relationship between the two modalities at a semantic level. These methods often suffer from high computational complexity and poor adaptability, making them difficult to apply to complex marine engineering environments. Summary of the Invention

[0003] Purpose of the invention: In response to the problems and shortcomings of the existing technology, the present invention proposes a semantic fusion detection method for underwater defects of hydraulic structures based on acoustic-optical imaging, which aims to build deep semantic associations between acoustic images and optical images through a deep learning model, and realize joint modeling and guided optimization of multimodal information. This method designs a semantic guidance mechanism so that the structural semantic information in the acoustic image can effectively guide the defect identification process in the optical image. While making full use of the stability of acoustic imaging, it integrates the texture details of the optical image to achieve high-precision recognition of multiple types of underwater defects such as cracks, erosion, and holes. At the same time, the fusion strategy proposed in the present invention can effectively control the complexity of the algorithm, improve the fusion efficiency, and has strong engineering practicality and promotion value.

[0004] Technical solution: A method for detecting underwater defects of hydraulic structures using acoustic-optical imaging semantic fusion, including: Step 1: Acquire optical image and sonar image data.

[0005] Step 2: Build an underwater image feature extraction model. The underwater image feature extraction model is used to extract multi-level, multi-dimensional visual information from optical images, including texture feature extraction, boundary feature extraction, and color feature extraction. The extracted features are then fused to obtain a unified image embedding feature representation.

[0006] Step 3: Construct an acoustic image semantic feature extraction model. The acoustic image semantic feature extraction model is used to extract structural semantic information from acoustic images and encode it into an embedding vector that can be used to guide image detection.

[0007] Step 4: Underwater acoustic-optical imaging semantic fusion detection model. This step constructs a multimodal defect detection network based on a semantic guidance mechanism, which aims to deeply fuse the visual information of the optical image with the semantic information of the acoustic image to improve the accuracy and robustness of detection. The optical image embedding features extracted in step 2 and the semantic vector obtained in step 3 are input into the parallel image encoder and text encoder respectively; the two encoders perform deep modeling on the input image region features and semantic category information respectively; the detection model realizes semantic-level region positioning by matching the correlation between the image region features and the semantic vector, that is, the sonar information guides the optical image to focus on the possible defect area for detection; the detection model also includes a bounding box prediction branch for accurately outputting the spatial position of the defect; and finally outputs the defect type, location information and detection confidence.

[0008] The design combines two imaging modalities: an optical imaging sensor and an acoustic imaging sensor. This ensures that the sensing areas of the acoustic and optical sensors overlap, thereby obtaining multimodal image data of the same structural area. The optical imaging sensor acquires optical data (optical images), while the acoustic imaging sensor acquires acoustic data (sonar images).

[0009] The underwater image feature extraction model includes the following three feature extraction branches: 1) Texture feature extraction unit: A neural network structure based on Gabor filtering is used, combined with a directional filter group and a convolutional feature encoder to extract local texture information from the image, including texture attributes such as frequency, direction, stripe structure and roughness, which is suitable for identifying defect areas with obvious texture changes such as cracks and erosion. 2) Boundary feature extraction unit: This module combines the Sobel operator with a deep edge detection network to detect and enhance the locations where structural mutations exist in the image, focusing on extracting information such as defect edges, contour mutations and crack boundaries, thereby improving the model's ability to recognize linear defects such as cracks and fractures. 3) Color feature extraction unit: The image is converted from RGB space to color space such as HSV or Lab, and features such as color distribution, brightness change and color offset of each channel are extracted to capture visual anomalies related to color changes such as corrosion and erosion. The above three feature extraction branches construct feature expressions of optical images from the three dimensions of texture, structure and color respectively. The output results are spliced ​​or weighted fused through a feature fusion module to finally generate a unified image embedding feature representation, denoted as F opt , used in the multimodal fusion process in subsequent semantic alignment and defect detection tasks.

[0010] The acoustic image semantic feature extraction model is used to extract structural semantic information from acoustic images and encode it into an embedding vector that can be used to guide image detection. It comprises: 1) Semantic Representation Encoding Unit: This unit first identifies defect categories in sonar images using a classification model based on a deep convolutional neural network to perform multi-category discrimination on the sonar images, identifying the presence and specific defect types, such as cracks, holes, and erosion. Based on the identified categories, a corresponding structured natural language description (e.g., "There is a crack," "There is a hole," etc.) is generated using pre-set mapping rules, achieving image-to-semantic translation. 2) Semantic Knowledge Recognition Unit: This unit inputs the natural language description as a semantic cue into the semantic knowledge recognition unit, where a pre-trained text encoding network (such as BERT or Transformer) is used to extract a semantic vector representation. This semantic vector carries prior knowledge about the defect category and is used to guide semantic detection in subsequent stages. Furthermore, the semantic feature extraction model can integrate a pre-set knowledge base or class embedding vectors to enhance the knowledge of the semantic representation and improve the generalization of semantic image understanding.

[0011] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned method for detecting underwater defects of hydraulic structures using acoustic-optical imaging semantic fusion are implemented.

[0012] A computer-readable storage medium stores a computer program for executing the above-mentioned acoustic-optical imaging semantic fusion detection method for underwater defects of hydraulic structures.

[0013] Beneficial effects: Based on the multimodal fusion flaw detection technology, the present invention develops and constructs a semantic guidance strategy, and innovatively proposes an acoustic-optical imaging semantic fusion detection framework, which effectively improves the operational efficiency of underwater defects of hydraulic structures while giving full play to the advantages of multimodal information fusion in terms of high accuracy in the process of underwater defect detection of hydraulic structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic flow diagram of the present invention; Figure 2 It is a schematic diagram of image feature extraction of the present invention; Figure 3 Schematic diagram of defect detection steps in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The present invention is further illustrated below with reference to specific examples. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0016] Existing underwater defect detection methods for hydraulic structures suffer from common problems such as insufficient detection accuracy, limited information acquisition, and susceptibility to interference from complex underwater environments. These issues make it difficult to meet the practical needs of efficient and accurate underwater defect detection. Furthermore, current multimodal detection technologies often face a conflict in practical applications, struggling to balance accuracy and operational efficiency, making it difficult to improve overall detection efficiency while maintaining accuracy.

[0017] To address the above issues, the present invention proposes an underwater defect detection method based on semantic-level acoustic-optical fusion. This method constructs an acoustic-optical semantic-level fusion framework to achieve deep fusion of acoustic and optical imaging information at the semantic feature level. This method fully leverages the advantages of acoustic imaging in terms of penetration, effectively overcoming the limitations of optical imaging in turbid water environments. At the same time, it introduces detailed information from the optical image to compensate for the insufficient resolution of the acoustic imaging, thereby improving the robustness, accuracy, and reliability of defect identification. Specifically, it includes the following: Step 1: By deploying the Blueprint Subsea Oculus M750d forward-looking sonar (FLS) underwater acoustic imaging sensor and the RGB underwater optical imaging sensor in the waters surrounding the hydraulic structure, the imaging areas of the acoustic sensor and the optical sensor overlap, and acoustic and optical images of the same hydraulic structure area are acquired simultaneously.

[0018] Step 2: The optical data (underwater optical image) acquired by the underwater optical imaging sensor in step 1 is input into the underwater image feature extraction model, which consists of a texture feature extraction unit, a boundary feature extraction unit, and a color feature extraction unit to extract the image features of the optical image. The specific network architecture is as follows: Figure 2As shown in the figure, the core goal of this network architecture is to extract semantic information through image classification. First, the network extracts features from the input image using multiple convolution-batch normalization-SiLU (CBS) modules, extracting image features at different levels. Then, the residual feature extraction module (C3K2) further refines the image's high-level semantic information. The network also includes the SPPF (Fast Spatial Pyramid Pooling) module, which uses multi-scale pooling to enhance the model's ability to capture features at different scales. The upsample module is used for upsampling, increasing the resolution of the feature map for better subsequent classification. Finally, the extracted features are classified by the classify module, and the classification results are further converted into semantic information. This semantic information is used to guide region selection and object detection in subsequent tasks.

[0019] The texture feature extraction unit uses a set of four directions and frequencies of Gabor filter kernels to convolve the image, and the expression is as follows: Among them F edge Represents the edge feature map, G x Represents the gradient of the image in the x direction, G y Indicates y Directional gradient; I gray A grayscale image representing an optical image.

[0020] The color feature extraction unit transforms the image from RGB to HSV and Lab space to extract color features. The expression is as follows:

[0021] Among them F opt Represents the fused multi-dimensional optical image features; W t ,W e ,W c It is the learned weight matrix, using different fully connected layers to map the three features to the same dimension, and then perform weighted fusion; Represents the feature weighting coefficient.

[0022] Step 3: Input the acoustic data (underwater sonar image) acquired by the underwater sonar imaging sensor in step 1 into the semantic feature extraction model. The acoustic image semantic feature extraction model consists of a semantic representation encoding unit and a semantic knowledge recognition unit to extract the semantic features of the acoustic image.

[0023] Semantic representation encoding unit: This unit first uses the YOLOV11 framework to extract and classify features of the image, then generates structured semantic text and performs embedded encoding.

[0024] Sonar image semantic classification uses a deep learning model to output the probability distribution of defect categories, expressed as follows:

[0025] Among them I sonar is the sonar image, f cls It is a classification model, P is the classification probability of different defects, and the category corresponding to the maximum probability is: Represents the defect label, which can be used to generate the corresponding semantic text description, p j Represents the first prediction of the acoustic image classification model j The probability value of class defect, j ∈{none, crack, hole, spalling}.

[0026] Based on the classification results, a fixed template is used to generate a structured semantic description, which is expressed as follows:

[0027]

[0028] Among them, TextGen() is a mapping function used to generate structured semantics.

[0029] The generated text T is input into the semantic encoder to extract the semantic vector representation, which is expressed as follows: t1=f text (T) where f text Represented as a text embedding network, t1 is the semantic embedding feature ultimately used to align with the optical image region. Semantic Knowledge Recognition Unit: This unit further semantically models the semantic vector and integrates the attention mechanism to improve semantic alignment capabilities. An external defect knowledge vector library is introduced to dynamically enhance the current semantic vector. The expression is as follows:

[0030] in , represents the attention weight of various types of knowledge; K j represents the knowledge vector of the j-th type of defect; t2 represents the enhanced semantic vector for subsequent cross-modal matching.

[0031] Step 4: Input the underwater optical image features and acoustic image semantic features obtained in steps 2 and 3 into the underwater acoustic-optical imaging semantic fusion detection model. The model uses a parallel Transformer-based image encoder and text encoder to process the image features of the underwater optical image and the semantic features of the acoustic image respectively. The specific model architecture is as follows: Figure 3The figure shows a network architecture for defect detection, which extracts semantic information from sonar images and combines them with optical images for defect detection. First, the network extracts semantic information (such as "no defect," "hole," etc.) from the sonar images using a semantic model and generates textual descriptions. These descriptions guide subsequent image processing. Next, the sonar and optical images undergo spatial and channel-wise convolution to extract multi-level features and improve feature representation. In the feature sharing module, features from the sonar and optical images are fused and dimensionality reduced using a pooling layer, using 3×3 and 1×1 convolutions to reduce computational complexity. These features are then fed into a fully connected layer for further processing, ultimately outputting a classification result to determine the defect category in the image. Finally, the network generates and displays the detection results, including a classification label and a localization bounding box that delineates the defect area and is displayed on the optical image, providing the operator with accurate defect location information. By combining features from sonar and optical images and leveraging the semantic information of the sonar image for guidance, this architecture significantly improves the accuracy and robustness of defect detection in complex environments.

[0032] The underwater acoustic-optical imaging semantic fusion detection model takes image-text pairs as input and is expressed as follows: (F opt ,T) Where T represents the category text semantic knowledge base, T={“There is no defect.”, “There is acrack.”, “There is a crack.”, “There is a spalling.”}.

[0033] The image encoder and text encoder encode the underwater optical image features and acoustic image semantic features respectively. The expressions are as follows:

[0034] Among them: F img Represents the image after encoding different features; is the image encoder's optical image texture feature F texture The encoding result is based on the texture feature F texture The obtained defect detection candidate area; is the image encoder's optical image boundary feature F edge The encoding result is based on the boundary feature F edge The obtained defect detection candidate area; is the image encoder's optical image boundary feature F color The encoding result is based on the boundary feature F colorThe obtained defect detection candidate area. Different features of semantics, It is the encoding result of the semantic vector t2 by the text encoder.

[0035] By calculating the underwater optical image feature encoding result F img and the acoustic image semantic feature encoding result F txt The cosine similarity of is used as the detection score S cls , the expression is as follows:

[0036] in ∈[0,1] represents the underwater optical image feature encoding result F img and the acoustic image semantic feature encoding result F txt degree of matching.

[0037] For each candidate region Using a separate box regressor for bounding box prediction, the expression is as follows:

[0038] in represents the predicted boundary coordinates of each candidate box, f reg represents the regressor, Represents the predicted value of the horizontal coordinate of the center point of the candidate box, Represents the predicted value of the vertical coordinate of the center point of the candidate box, Represents the predicted value of the width of the candidate box, Indicates the predicted value of the candidate box's height.

[0039] The final model is optimized through a joint loss, including semantic matching loss and box positioning loss.

[0040] The semantic alignment loss expression is as follows:

[0041] in represents the cross entropy loss for matching image regions with text categories, y ij ∈{0,1} represents the ground truth matching label of the region-category.

[0042] The box localization loss expression is as follows:

[0043] in It is the positioning loss of the bounding box, smoothing the L1 loss function, and calculating the position error between the predicted bounding box and the real box; represents the real frame, Represents the predicted box.

[0044] The total loss expression is as follows:

[0045] in Represents the loss weighting factor.

[0046] Obviously, those skilled in the art should understand that the various steps of the acoustic-optical imaging semantic fusion detection method for underwater defects in hydraulic structures of the above-mentioned embodiment of the present invention can be implemented using a general-purpose computing device. They can be concentrated on a single computing device or distributed on a network composed of multiple computing devices. Optionally, they can be implemented using program codes executable by the computing device, so that they can be stored in a storage device and executed by the computing device. In some cases, the steps shown or described can be performed in a different order than herein, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the embodiments of the present invention are not limited to any specific combination of hardware and software.

Claims

1. A method for detecting underwater defects of hydraulic structures by using acoustic-optical imaging semantic fusion, characterized in that: include: Synchronously acquire underwater optical image and sonar image data of the same hydraulic structure area; Based on the underwater image feature extraction model, the image features of the underwater optical image are extracted, including the texture, boundary and color features of the underwater optical image, and the texture, boundary and color features are fused into image features; the acoustic image semantic feature extraction model is used to extract structural semantic information from the acoustic image and encode it into an embedded semantic vector that can be used to guide image detection; the image features and semantic vector are input in parallel into the underwater acoustic-optical imaging semantic fusion detection model, and the image features of the optical image are processed under the guidance of the acoustic image semantic features to output the underwater defect detection results.

2. The method for detecting underwater defects of hydraulic structures by acoustic-optical imaging semantic fusion according to claim 1, characterized in that: The underwater image feature extraction model includes the following three feature extraction branches: 1) Texture feature extraction unit: a neural network structure based on Gabor filtering is used, combined with a directional filter bank and a convolutional feature encoder to extract local texture information from the image, including the frequency, direction, stripe structure and roughness of the texture; 2) Boundary feature extraction unit: a Sobel operator is combined with a deep edge detection network to detect and enhance the locations where structural mutations exist in the image, and extract defect edges, contour mutations and crack boundary information; 3) Color feature extraction unit: the image is converted from RGB space to HSV or Lab color space, and the color distribution, brightness change and color shift features of each channel are extracted to capture visual anomalies related to corrosion, erosion and color change; the above three feature extraction branches construct feature expressions of optical images from the three dimensions of texture, structure and color respectively, and their output results are spliced ​​or weighted fused through a feature fusion module to finally generate a unified image embedding feature representation, which is recorded as image feature F opt , used for the multimodal fusion process in subsequent semantic alignment and defect detection tasks.

3. The method for detecting underwater defects of hydraulic structures by acoustic-optical imaging semantic fusion according to claim 1, characterized in that: The device for acquiring underwater optical images and sonar image data of the same hydraulic structure area is an imaging device that combines two modalities. The first is an optical imaging sensor, and the second is an acoustic imaging sensor. The optical imaging sensor device is used to acquire optical image data, and the acoustic imaging sensor device is used to acquire sonar image data. The sensing areas of the acoustic sensor and the optical sensor have spatial overlap, thereby obtaining multimodal image data of the same structural area.

4. The method for detecting underwater defects of hydraulic structures by acoustic-optical imaging semantic fusion according to claim 1, characterized in that: The underwater acoustic-optical imaging semantic fusion detection model is constructed based on a multimodal defect detection network based on a semantic guidance mechanism. It aims to fuse the visual information of the optical image with the semantic information of the acoustic image, and input the optical image features and semantic vectors into a parallel image encoder and text encoder respectively. The two encoders respectively model the input image region features and semantic category information. The detection model achieves semantic-level region positioning by matching the degree of correlation between the image region features and the semantic vectors, that is, the optical image is guided by sonar information to detect possible defect areas. The detection model also includes a bounding box prediction branch for outputting the spatial location of the defect. Finally, the defect type, location information and detection confidence are output.

5. The method for detecting underwater defects of hydraulic structures by acoustic-optical imaging semantic fusion according to claim 1, characterized in that: The acoustic image semantic feature extraction model is used to extract structural semantic information from acoustic images and encode it into an embedded vector that can be used for image detection guidance. It includes: 1) a semantic representation encoding unit: first, the sonar image is defect-classified and a classification model based on a deep convolutional neural network is used to perform multi-category discrimination on the sonar image to identify whether there are defects and the specific type of defects; then, based on the identified categories, a corresponding structured natural language description is generated through preset mapping rules to achieve the conversion from image to semantics; 2) a semantic knowledge recognition unit: the natural language description is input into the semantic knowledge recognition unit as a semantic prompt, and a semantic vector representation is extracted through a pre-trained text encoding network; the semantic vector carries prior knowledge about the defect category and is used for semantic guidance of the image detection area in the subsequent stage.

6. The method for detecting underwater defects of hydraulic structures by acoustic-optical imaging semantic fusion according to claim 2, characterized in that: The texture feature extraction unit uses a set of four-direction and frequency Gabor filter kernels to convolve the underwater optical image; the color feature extraction unit transforms the image from RGB to HSV and Lab space to extract color features, as shown in the following expression: in, represents the color feature map, Represents vector concatenation operation, Hist represents color histogram function; Represents an optimized RGB image; Represents an optimized HSV image; Represents an optimized Lab image; The texture features, boundary features and color features are integrated as follows: ， Among them F opt Represents the fused multi-dimensional optical image features; It is the learned weight matrix, using different fully connected layers to map the three features to the same dimension, and then perform weighted fusion; Represents the feature weighting coefficient.

7. The method for detecting underwater defects of hydraulic structures by acoustic-optical imaging semantic fusion according to claim 5, characterized in that: In the acoustic image semantic feature extraction model, the semantic representation encoding unit first extracts and classifies the image features using the YOLOV11 framework, then generates structured semantic text and performs embedded encoding. For sonar image semantic classification, a deep learning model is used to output the probability distribution of defect categories. The category corresponding to the maximum probability is: Represents the defect label, which is used to generate the corresponding semantic text description, p j Represents the first prediction of the acoustic image classification model j The probability value of class defect, j ∈{none,crack,hole,spalling}; Based on the classification results, a fixed template is used to generate a structured semantic description to obtain the generated text T, which is then input into the semantic encoder to extract the semantic vector representation; Semantic knowledge recognition unit: This unit further models the semantic vector and integrates the attention mechanism to improve semantic alignment capabilities. It also introduces an external defect knowledge vector library to dynamically enhance the current semantic vector. The expression is as follows: in Represents the attention weights of various types of knowledge; represents the knowledge vector of the j-th type of defect; t2 represents the enhanced semantic vector for subsequent cross-modal matching, and t1 is the semantic embedding feature ultimately used for alignment with the optical image region.

8. The method for detecting underwater defects of hydraulic structures by acoustic-optical imaging semantic fusion according to claim 1, characterized in that: In the underwater acoustic-optical imaging semantic fusion detection model, the underwater optical image feature encoding result F is calculated. img and the acoustic image semantic feature encoding result F txt The cosine similarity of is used as the detection score S cls , the expression is as follows: in ∈[0,1] represents the underwater optical image feature encoding result F img and the acoustic image semantic feature encoding result F txt degree of matching; For each candidate region Using an independent box regressor for bounding box prediction, the expression is as follows: in represents the predicted boundary coordinates of each candidate box, represents a regressor; Represents the predicted value of the horizontal coordinate of the center point of the candidate box, Represents the predicted value of the vertical coordinate of the center point of the candidate box, Represents the predicted value of the width of the candidate box, Represents the predicted value of the height of the candidate box; The final model is optimized through a joint loss, including semantic matching loss and box positioning loss; The semantic alignment loss expression is as follows: in represents the cross entropy loss for matching image regions with text categories, Represents the ground truth matching label of region-category; The box localization loss expression is as follows: in is the bounding box localization loss; The total loss expression is as follows: in Represents the loss weighting factor.

9. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the acoustic-optical imaging semantic fusion detection method for underwater defects of hydraulic structures as described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program for executing the acoustic-optical imaging semantic fusion detection method for underwater defects of hydraulic structures according to any one of claims 1 to 8.

Images