A Smart Underwater Crack Detection and Size Quantization Model

Abstract

This invention relates to the field of image processing technology and discloses an intelligent underwater crack detection and size quantization model, including a preprocessing module, an underwater crack structure perception and segmentation network, and a crack size quantization module. The underwater crack structure perception and segmentation network includes an encoder, a multi-scale context enhancement module, a decoder with CSRM, and a crack segmentation head. The preprocessing module receives the original underwater dam surface image and generates an image to be detected. The underwater crack structure perception and segmentation network receives the image to be detected and obtains a crack binary segmentation mask. The crack size quantization module receives the crack binary segmentation mask and generates a visualized image of underwater crack identification and size quantization. This invention, by combining the underwater crack structure perception and segmentation network and the crack size quantization module, achieves automatic crack identification, structural enhancement, and size quantization in complex underwater environments, providing more reliable technical support for underwater structural defect detection, damage assessment, and maintenance decisions.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to an intelligent underwater crack detection and size quantification model. Background Technology

[0002] Underwater structures are widely found in dams, bridge piers, port terminals, tunnels, sluice gates, water pipelines, and marine engineering structures. Due to long-term exposure to water flow erosion, pressure changes, chemical corrosion, temperature variations, and external loads, defects such as cracks, erosion, and voids easily develop on the structural surface. Among these, cracks are an important type of defect reflecting the structural safety status, and their length, width, area, and development trend are often used as important bases for structural safety assessments and maintenance decisions.

[0003] Traditional underwater crack detection mainly relies on manual diving inspections or image acquisition by remotely operated underwater robots followed by manual interpretation. This method suffers from low detection efficiency, high subjectivity, high risk, and difficulty in quantifying crack size. In recent years, semantic segmentation methods based on deep learning have been used for crack identification, but underwater crack images differ significantly from ordinary ground crack images. Underwater imaging is often affected by light attenuation, water turbidity, suspended particles, reflections, water ripples, local blurring, and background texture interference, causing cracks to exhibit characteristics such as low contrast, blurred boundaries, local discontinuity, and significant scale variations.

[0004] Existing semantic segmentation methods typically focus on improving pixel classification accuracy, but pay insufficient attention to the continuity of the elongated structure and the quantification of the size of cracks. On the one hand, ordinary convolutional or feature fusion structures can easily submerge the features of elongated cracks in complex backgrounds; on the other hand, existing methods mostly output binary segmentation results, making it difficult to directly provide quantitative indicators such as crack length, width, and area required for engineering detection.

[0005] For example, patent document CN121524553A discloses a method for detecting underwater surface cracks in hydraulic concrete structures, which relates to the field of image sensor technology. The method includes: mapping the underwater light field characterization parameters to transmittance and separating the background radiation field to generate a target reflectance image; fusing multispectral-polarization data in the target reflectance image to generate an enhanced feature image; performing multi-scale convolutional kernel spatial transformation and feature pyramid fusion on the enhanced feature image to generate a crack response map; performing anisotropic nonmaximum suppression and morphological skeleton refinement on the crack response map to generate crack detection data; and performing geometric topology joint optimization and parameter quantization on the crack detection data to output a structured detection report.

[0006] Therefore, it is necessary to propose a technical solution for the identification and dimensional quantification of surface cracks on underwater structures, which can improve the accuracy of crack identification and realize the automatic calculation and visual annotation of crack geometry. Summary of the Invention

[0007] To address the technical problems of existing technologies in identifying slender crack features in complex backgrounds and in quantifying the size of the identified cracks, this invention provides an intelligent underwater crack detection and size quantization model, comprising: a preprocessing module, an underwater crack structure perception and segmentation network, and a crack size quantization module; the underwater crack structure perception and segmentation network comprises: an encoder, a multi-scale context enhancement module, a decoder with CSRM, and a crack segmentation head; The preprocessing module is used to receive the original underwater dam surface image, perform preprocessing, generate the image to be detected, and transmit it to the encoder. The encoder is used to receive the image to be detected, extract crack features at different scales step by step to obtain multi-level features; the multi-level features include shallow crack detail features, mid-level crack structure features and deep semantic features; the deep semantic features are input to the multi-scale context enhancement module; the shallow crack detail features and mid-level crack structure features are input to the decoder with CSRM. The multi-scale context enhancement module is used to receive the deep semantic features, perform multi-scale context residual enhancement, obtain the deep semantic features after residual enhancement, and transmit them to the decoder with CSRM. The decoder with CSRM is used to receive the shallow crack detail features, the mid-level crack structure features, and the residual-enhanced deep semantic features, perform feature fusion, and obtain multi-level fused features. The crack structure perception and refinement module CSRM generates a crack structure response map and refined crack features. The refined crack features Transmitted to the crack splitting head; The crack segmentation head is used to receive the refined crack features. Output crack probability map Binarization is performed to obtain the crack binary segmentation mask M, which is then transmitted to the crack size quantization module CSQM. The crack size quantization module CSQM is used to receive the crack binary segmentation mask M, calculate the area, length and width of the crack binary segmentation mask M, and generate a visual image of underwater crack identification and size quantization; the visual image includes crack identification results and size annotation information.

[0008] Furthermore, the preprocessing includes format unification and size unification; the unified format supports RGB format, and the unified size supports 512×512 pixels.

[0009] Furthermore, the encoder includes a stem layer and four feature extraction stages; the stepwise extraction of crack features at different scales includes: The image to be detected is passed through a stem layer to obtain the initial shallow crack detail features F1; For the initial shallow crack detail feature F1, an improved LASS module and convolutional downsampling structure are used to sequentially obtain features F2, F3, F4 and F5 through four feature extraction stages. Among them, feature F2 is the shallow crack detail feature, features F3 and F4 are the mid-level crack structure features, and feature F5 is the deep semantic feature; the shallow crack detail feature includes crack edge, texture and local detail information; the mid-level crack structure feature includes crack morphology and local structure information; and the deep semantic feature includes semantic context information.

[0010] Furthermore, the multi-scale context enhancement module includes multiple parallel dilated convolution branches; the multi-scale context residual enhancement includes: Based on the deep semantic features, different dilation rates r are set for the multiple parallel dilated convolution branches, context modeling is performed, and branch features at different receptive field scales are output. ; The branch features at different receptive field scales are concatenated along the channel dimension and fused using 1×1 convolution to obtain multi-scale contextual residual enhancement features. ; The multi-scale contextual residual enhancement feature The deep semantic features are fused with the original deep semantic features to obtain the residual-enhanced deep semantic features, thus completing the multi-scale contextual residual enhancement.

[0011] Furthermore, the acquisition of multi-level fusion features include: The shallow crack detail features are used as the reference features; The mid-level crack structure features and the residual-enhanced deep semantic features are concatenated with the baseline features through channel mapping and upsampling to unify the channel dimension and spatial scale, thereby obtaining multi-level fused features. .

[0012] Furthermore, the generated crack structure response map and refined crack features include: The multi-level fusion features The input is a crack structure perception refinement module (CSRM), and the structure response head of the CSRM is used to refine the multi-level fused features. Mapping is performed to obtain the crack structure response map; The crack structure response map is combined with the multi-level fusion feature. Element-wise multiplication yields the structural enhancement features. Perform 3×3 thinned convolution and fuse it with the multi-level features. By summing the residuals, the refined crack features are obtained. .

[0013] Furthermore, the crack probability map The expression is: ; in, H and W represent the spatial height and width of the image to be recognized, respectively; This is the Sigmoid activation function.

[0014] Furthermore, obtaining the binary segmentation mask M for the crack includes: Set a threshold T, and use the threshold T to analyze the crack probability map. Binarization is performed to obtain the binary segmentation mask M for the crack, which is expressed as: ; Where P(i,j) represents the probability map of pixel (i,j) in the crack. The probability value above.

[0015] Furthermore, the calculation of the area, length, and width of the binary segmentation mask M for the crack includes: Perform connected component analysis on the binary segmentation mask M of the crack, and count the number of connected regions of the crack. Calculate the pixel area of ​​the crack ; Obtain the actual length y corresponding to a single pixel, and calculate the actual area of ​​the crack. ; The binary segmentation mask M of the crack is skeletonized to obtain the crack centerline S and the number of skeleton pixels, and the crack length is calculated. Obtain the pixel calibration ratio z and calculate the actual length of the crack. ; Perform a distance transform on the binary segmentation mask M of the crack to obtain the distance from each pixel within the crack region to the nearest background pixel. Calculate the average width of local cracks and the maximum width of local cracks Based on the pixel calibration ratio z, the actual average width of the local crack is calculated. and the actual maximum width of local cracks .

[0016] Furthermore, the generation of the visualized image for underwater crack identification and size quantification includes: The actual area of ​​the crack Actual length of the crack The actual average width of local cracks and the actual maximum width of local cracks The data is statistically analyzed to generate a comprehensive dimension report. The dimension data in the comprehensive dimension report is then overlaid onto the original underwater dam surface image to generate a visual image for underwater crack identification and dimension quantification.

[0017] The beneficial effects of this invention are: (1) This invention introduces the Lite-ASPP multi-scale context enhancement module into the deep features of the UCSNet network. It extracts contextual information under different receptive fields through parallel convolutional branches with different dilation rates, and forms enhanced deep semantic features through feature concatenation and 1×1 convolution fusion. It can enhance the model's adaptability to cracks of different widths, lengths and weak contrasts while keeping the computational load relatively controllable; reduce crack missed detection and fracture phenomenon caused by insufficient receptive field; and improve the overall recognition ability of crack regions in complex underwater backgrounds.

[0018] (2) This invention introduces a crack structure perception and refinement module (CSRM) into the decoder. Using multi-level fusion features as input, a crack structure response map is generated through a structure response head. This response map is then used to weight the fusion features element-wise, enhancing the effective responses related to the main crack structure, the central extension region, and the weak boundary region. The enhanced structure features are then refined through convolution and added back to the original fusion features as residuals to obtain the refined crack features. The CSRM used in this invention can improve the network's ability to express slender cracks, discontinuous cracks, and weak boundary cracks without increasing the cost of additional skeleton annotation and additional supervision constraints, making the prediction results more continuous and complete. Simultaneously, the structure response weighting can suppress underwater background texture and noise interference, reducing false detections and false negatives, and improving the structural consistency and boundary clarity of the crack segmentation results.

[0019] (3) This invention sets up a crack size quantification module (CSQM) after crack segmentation prediction. Taking the crack binary segmentation mask as input, it sequentially performs connected component analysis, skeleton extraction, distance transformation, and size statistics. The number of crack regions is counted and the crack area is calculated through connected component analysis; the crack centerline is obtained through skeleton extraction, and the crack length is estimated based on the number of skeleton pixels; the distance from the pixel to the boundary within the crack region is calculated through distance transformation, and the local crack width is estimated at the skeleton point to obtain the average width and maximum width. When the pixel physical calibration ratio is known, the pixel-level size can also be converted into the actual physical size. CSQM can further transform the crack segmentation results into quantitative indicators that can be used for engineering evaluation, realize the integrated processing of crack identification, skeleton analysis, width estimation, and size statistics, automatically generate information such as crack area, length, average width, maximum width, and number of connected components, and overlay the crack identification results and size annotations onto the original underwater image to form an intuitive visualization output, improving the interpretability of the detection results and the engineering application value.

[0020] (4) This invention combines the UCSNet crack segmentation network with the CSQM crack size quantization module to form a complete process from underwater image input, crack region identification, structure-aware refinement, binary mask generation to size quantization output. The front-end UCSNet is responsible for obtaining relatively complete and continuous crack segmentation results in complex underwater environments; the back-end CSQM performs skeleton extraction, distance transformation, and size statistics based on the segmentation results, and outputs engineering-readable quantization results. Through the collaborative work of the two, this invention can achieve automatic crack identification, structure enhancement, and size quantization under conditions such as low contrast, uneven illumination, and complex background texture underwater, providing more reliable technical support for underwater structural defect detection, damage assessment, and maintenance decisions. Attached Figure Description

[0021] Figure 1 This is a structural diagram of the intelligent underwater crack detection and size quantification model provided by the present invention; Figure 2 This is a diagram showing the UCSNet crack segmentation prediction effect provided by the present invention; Figure 3 This is a visualization result of CSQM crack size quantification provided by the present invention. Detailed Implementation

[0022] The technical solution of the present invention is further described below, but the scope of protection is not limited to what is described.

[0023] This invention provides an intelligent underwater crack detection and size quantification model, the model structure of which is as follows: Figure 1As shown, it includes: a preprocessing module, an underwater crack structure perception and segmentation network (UCSNet), and a crack size quantization module (CSQM); the underwater crack structure perception and segmentation network (UCSNet) includes: an encoder, a multi-scale context enhancement module (Lite-ASPP), a decoder with CSRM, and a crack segmentation head. The preprocessing module is used to receive the original underwater dam surface image, perform preprocessing, generate the image to be detected, and transmit it to the encoder; the original underwater dam surface image includes the original image of the underwater dam body, gate, pier, or other underwater concrete structure surface.

[0024] The preprocessing includes format and size standardization; the standardized format supports RGB format, and the standardized size supports 512×512 pixels. For the crack annotation map used in the training phase of the intelligent underwater crack detection and size quantization model, it is converted into a single-channel grayscale mask and binarized, so that the pixel value of the crack area is 1 and the pixel value of the background area is 0.

[0025] The dimensions were standardized using a normalization process to reduce the impact of uneven underwater lighting, turbidity variations, and differences in imaging equipment on model training and inference. During the training phase, online data augmentation techniques such as random rotation, horizontal flipping, vertical flipping, and slight contrast enhancement were employed to improve the model's robustness to different shooting postures, crack directions, and background texture interference.

[0026] The underwater crack structure perception and segmentation network is UCSNet, which adopts an encoder-decoder structure.

[0027] The encoder is used to receive the image to be detected, extract crack features at different scales step by step to obtain multi-level features; the multi-level features include shallow crack detail features, mid-level crack structure features and deep semantic features; the deep semantic features are input into the multi-scale context enhancement module (Lite-ASPP); the shallow crack detail features and mid-level crack structure features are input into the decoder with CSRM. The encoder includes a stem layer and four feature extraction stages for progressively extracting crack features at different scales; the progressive extraction of crack features at different scales includes: The image to be detected is passed through a stem layer to obtain the initial shallow crack detail features F1; For the initial shallow crack detail feature F1, an improved LASS module and convolutional downsampling structure are used to sequentially obtain features F2, F3, F4 and F5 through four feature extraction stages. Among them, feature F2 is the shallow crack detail feature, features F3 and F4 are the mid-level crack structure features, and feature F5 is the deep semantic feature; the shallow crack detail feature includes crack edge, texture and local detail information; the mid-level crack structure feature includes crack morphology and local structure information; and the deep semantic feature includes semantic context information.

[0028] Specifically, let the image to be detected be X. After passing through the encoder, multi-level features F1, F2, F3, F4, and F5 are obtained. Among them, shallow features F1 and F2 contain crack edges, texture, and local detail information; middle features F3 and F4 contain crack morphology and local structural information; and deep feature F5 contains semantic context information.

[0029] During the encoding process, an improved LASS module and convolutional downsampling structure are used to extract features stepwise from the slender structure, weak boundary regions, and locally discontinuous regions of the underwater crack. Through multi-stage encoding, the underwater crack structure-aware segmentation network (UCSNet) can simultaneously preserve shallow detail information and deep semantic information, providing a foundation for subsequent multi-scale context enhancement and structure-aware refinement.

[0030] The multi-scale context enhancement module (Lite-ASPP) is used to receive the deep semantic features, enhance the multi-scale contextual expression capability, obtain the enhanced deep semantic features, and transmit them to the decoder with CSRM. Due to the significant scale differences, elongated shape, blurred boundaries, and local discontinuity of underwater cracks, it is difficult to reliably identify cracks of different widths and lengths using only a single receptive field. Therefore, a lightweight void spatial pyramid pooling (Lite-ASPP multi-scale context enhancement module) is introduced at the deep feature F5 to enhance the multi-scale contextual representation capability of the deep feature.

[0031] The multi-scale context enhancement module (Lite-ASPP) includes multiple parallel dilated convolution branches; the multi-scale contextual representation enhancement includes: Based on the deep semantic features, different dilation rates r are set for the multiple parallel dilated convolution branches, context modeling is performed, and branch features at different receptive field scales are output. ; The branch features at different receptive field scales are concatenated along the channel dimension and fused using 1×1 convolution to obtain multi-scale context-enhanced features. ; The multi-scale context enhancement features By fusing with the original deep semantic features, enhanced deep semantic features are obtained, thus completing the enhancement of multi-scale contextual expression capabilities.

[0032] In this embodiment, the void ratio r is set to 1, 3, 6, and 9. Let the branch output for void ratio r be: (1) Where d=r represents the hole rate of the convolution branch.

[0033] This yields four branch features at different receptive field scales: (2) (3) (4) (5) Among them, the branch with a void ratio of 1 is used to preserve local semantic information; the branches with void ratios of 3, 6 and 9 are used to expand the receptive field and capture the contextual relationship between cracks at different scales and their surrounding background.

[0034] The outputs of each branch are concatenated along the channel dimension and then fused using a 1×1 convolution to obtain multi-scale contextual residual enhancement features. : (6) To preserve the original deep semantic information, residual enhancement features are used. By fusing with the original deep features, we obtain the residual-enhanced deep semantic features: (7) With the Lite-ASPP module, the Underwater Crack Structure Sensing and Segmentation Network (UCSNet) can enhance the ability to detect cracks of different widths, lengths, and low contrasts with relatively controllable computational cost, reducing the occurrence of missed cracks and fractures due to insufficient single receptive field.

[0035] The decoder with CSRM is used to receive the shallow crack detail features, the mid-level crack structure features, and the residual-enhanced deep semantic features, perform feature fusion, and obtain multi-level fused features. The crack structure perception and refinement module CSRM generates a crack structure response map and refined crack features. The refined crack features Transmitted to the crack splitting head; The decoder with CSRM receives multi-level features from the encoder (shallow crack detail features F2, mid-level crack structure features F3 and F4) and deep semantic features enhanced by Lite-ASPP residuals.

[0036] The acquisition of multi-level fusion features include: The shallow crack detail features are used as the reference features; The mid-level crack structure features and the residual-enhanced deep semantic features are concatenated with the baseline features through channel mapping and upsampling to unify the channel dimension and spatial scale, thereby obtaining multi-level fused features. The channel mapping is achieved through 1×1 convolution.

[0037] Specifically, for example: let the multi-level features after unifying the spatial scale be as follows: (8) (9) (10) (11) Align represents an operation that unifies the spatial scale, such as upsampling.

[0038] By concatenating the above features, a multi-level fused feature is obtained. : (12) The multi-level fusion features It also includes crack edge details, local textures, structural morphology, and global context information, which can be used for subsequent crack structure response modeling and structure-aware refinement.

[0039] The decoder with CSRM introduces a crack structure perception and refinement module (CSRM); the CSRM uses multi-level feature fusion. As input, a crack structure response map is generated through the structure response header. This crack structure response map originates from the features of the Underwater Crack Structure Aware Segmentation Network (UCSNet) itself, and is used to represent the features fused from the multi-level features by the UCSNet network. The crack structure sensitive area learned in the process.

[0040] In this embodiment, CSRM is a feature-level structure-aware refinement mechanism that does not introduce additional skeleton labels, does not set up an independent skeleton supervision branch, and does not write the skeleton extraction results into the training loss function. The UCSNet network still performs end-to-end optimization based on the final crack segmentation target.

[0041] The generated crack structure response map and refined crack features include: The multi-level fusion features The input is a crack structure perception refinement module (CSRM), and the structure response head of the CSRM is used to refine the multi-level fused features. Mapping is performed to obtain the crack structure response map R; the expression for the crack structure response map R is: (13) Where σ represents the Sigmoid activation function, and R represents the crack structure response map.

[0042] The crack structure response map is combined with the multi-level fusion feature. Element-wise multiplication yields the structural enhancement features. The structural enhancement features The expression is: (14) Where ⊙ represents element-wise multiplication. The structural enhancement feature... The acquisition process can enhance the response associated with the crack center structure, slender extension regions and weak boundary regions, while suppressing interference from underwater background textures, water stains, shadows and noise.

[0043] Perform 3×3 thinned convolution and fuse it with the multi-level features. By summing the residuals, the refined crack features are obtained. The refined crack features The expression is: (15) (16) in, This represents a 3×3 thinning convolution operation. This represents the residual feature obtained after 3×3 thinning convolution.

[0044] By using residual refinement, CSRM can highlight the crack structure response while preserving the original multi-level fusion information, and enhance the ability to express slender cracks, weak boundary cracks and local discontinuous cracks.

[0045] The crack segmentation head is used to receive the refined crack features. Output crack probability map Binarization is performed to obtain the crack binary segmentation mask M, which is then transmitted to the crack size quantization module CSQM. The crack probability map The expression is: (17) in, H and W represent the spatial height and width of the image to be identified, respectively. In this embodiment, they are obtained by image preprocessing and normalization, and both have a value of 512. This is the Sigmoid activation function.

[0046] The process of obtaining the binary segmentation mask M for the crack includes: Set a threshold T, and use the threshold T to analyze the crack probability map. Binarization is performed to obtain the binary segmentation mask M for the crack, which is expressed as: (18) Where P(i,j) represents the probability map of pixel (i,j) in the crack. The probability value on; Crack probability map The pixel coordinates of the i-th row and j-th column, i , T is the segmentation threshold, and in this embodiment, T=0.5.

[0047] Since crack pixels typically account for a small percentage of the overall underwater structure surface image, the model training employs an IoU loss function suitable for scenes with sparse foreground regions. Let the predicted probability map be P, and the ground truth crack annotation mask be G, then the soft IoU is defined as: (19) The corresponding IoU loss function is: (20) Here, k represents the pixel index, and ε is a smoothing coefficient used to avoid a denominator of zero. The above loss directly affects the final crack segmentation result, causing the UCSNet network to optimize around the overall overlap between the crack region and the ground truth labeled region.

[0048] The crack size quantization module CSQM receives the crack binary segmentation mask M, calculates its area (i.e., binary crack mask acquisition), length (i.e., skeleton extraction), and width (i.e., distance transformation) to generate a visual image of underwater crack identification and size quantization. The visual image includes crack identification results and size annotation information. The crack size quantization module CSQM outputs indicators such as crack area, length, average width, maximum width, and number of connected components.

[0049] The calculation of the area, length, and width of the binary segmentation mask M for the crack includes: Perform connected component analysis on the binary segmentation mask M of the crack, and count the number of connected regions of the crack. The number of connected crack regions is statistically significant. The expression is: (twenty one) Calculate the pixel area of ​​the crack The expression is: (twenty two) Obtain the actual length y corresponding to a single pixel, and calculate the actual area of ​​the crack. The expression is: (twenty three) The binary segmentation mask M of the crack is skeletonized to obtain the crack centerline S and the number of skeleton pixels. The expression for the crack centerline S is: (twenty four) Calculate crack length The expression is: (25) Obtain the pixel calibration ratio z and calculate the actual length of the crack. The expression is: (26) Perform a distance transform on the binary segmentation mask M of the crack to obtain the distance from each pixel within the crack region to the nearest background pixel. , The expression is: (27) Where p represents the pixel within the crack region, q represents the pixel within the background region, and Ω bg This represents the set of background pixels.

[0050] For a skeleton point p∈S, its local crack width can be approximated as: (28) Therefore, the average width of local cracks is calculated. and the maximum width of local cracks The expression is: (29) (30) Based on the pixel calibration ratio z, calculate the actual average width of the local crack. and the actual maximum width of local cracks The expression is: (31) (32) The generated visualization images for underwater crack identification and size quantification include: The actual area of ​​the crack Actual length of the crack The actual average width of local cracks and the actual maximum width of local cracks The data is statistically analyzed (dimension quantification) to generate a comprehensive dimension report (dimension quantification report). The dimension data in the comprehensive dimension report is then overlaid onto the original underwater dam surface image to generate a visual image for underwater crack identification and dimension quantification.

[0051] like Figure 2 and Figure 3 As shown, where, Figure 2 This paper demonstrates a pixel-by-pixel comparison between the crack prediction mask M (UCSNet) output by UCSNet and the manually annotated map (annotation) on four different low-contrast underwater real dam samples (a, b, c, and d). Figure 2 The three rows from top to bottom are the original image, the manually annotated image (annotation), and the crack prediction mask M (UCSNet). Figure 3 This displays the quantization report image automatically generated by CSQM. The generated visualization image in the left panel uses a red semi-transparent overlay to represent the crack binary prediction mask M, and uses a golden outer rectangle to dynamically select and label the number of each connected region. The data, including the crack's physical area A, estimated length L, and maximum width, is displayed in a table format in a separate data panel on the right. Average width and the total number of connected components The system incorporates a complete set of defect feature parameters (comprehensive dimensional report data), thereby establishing an integrated mapping from image-level geometric topology transformation to engineering-level dimensional statistics, which can be directly used for the detection, recording, and defect safety assessment of underwater structure surfaces.

[0052] The above-disclosed embodiments are merely specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. An intelligent underwater crack detection and size quantification model, characterized in that, include: Preprocessing module, underwater crack structure perception and segmentation network, and crack size quantization module; The underwater crack structure perception and segmentation network includes: an encoder, a multi-scale context enhancement module, a decoder with CSRM, and a crack segmentation head; The preprocessing module is used to receive the original underwater dam surface image, perform preprocessing, generate the image to be detected, and transmit it to the encoder. The encoder is used to receive the image to be detected, extract crack features at different scales step by step to obtain multi-level features; the multi-level features include shallow crack detail features, mid-level crack structure features and deep semantic features; the deep semantic features are input to the multi-scale context enhancement module; the shallow crack detail features and mid-level crack structure features are input to the decoder with CSRM. The multi-scale context enhancement module is used to receive the deep semantic features, perform multi-scale context residual enhancement, obtain the deep semantic features after residual enhancement, and transmit them to the decoder with CSRM. The decoder with CSRM is used to receive the shallow crack detail features, the mid-level crack structure features, and the residual-enhanced deep semantic features, perform feature fusion, and obtain multi-level fused features. The crack structure perception and refinement module CSRM generates a crack structure response map and refined crack features. The refined crack features Transmitted to the crack splitting head; The crack segmentation head is used to receive the refined crack features. Output crack probability map Binarization is performed to obtain the crack binary segmentation mask M, which is then transmitted to the crack size quantization module CSQM. The crack size quantization module CSQM is used to receive the crack binary segmentation mask M, calculate the area, length and width of the crack binary segmentation mask M, and generate a visual image of underwater crack identification and size quantization; the visual image includes crack identification results and size annotation information.

2. The intelligent underwater crack detection and size quantification model as described in claim 1, characterized in that, The preprocessing includes format unification and size unification; the format unification supports RGB format, and the size unification supports 512×512 pixels.

3. The intelligent underwater crack detection and size quantification model as described in claim 1, characterized in that, The encoder includes a stem layer and four feature extraction stages; the stepwise extraction of crack features at different scales includes: The image to be detected is passed through a stem layer to obtain the initial shallow crack detail features F1; For the initial shallow crack detail feature F1, an improved LASS module and convolutional downsampling structure are used to sequentially obtain features F2, F3, F4 and F5 through four feature extraction stages. Among them, feature F2 is the shallow crack detail feature, features F3 and F4 are the mid-level crack structure features, and feature F5 is the deep semantic feature; the shallow crack detail feature includes crack edge, texture and local detail information; the mid-level crack structure feature includes crack morphology and local structure information; and the deep semantic feature includes semantic context information.

4. The intelligent underwater crack detection and size quantification model as described in claim 1, characterized in that, The multi-scale context enhancement module includes multiple parallel dilated convolution branches; The multi-scale contextual residual enhancement includes: Based on the deep semantic features, different dilation rates r are set for the multiple parallel dilated convolution branches, context modeling is performed, and branch features at different receptive field scales are output. ; The branch features at different receptive field scales are concatenated along the channel dimension and fused using 1×1 convolution to obtain multi-scale contextual residual enhancement features. ; The multi-scale contextual residual enhancement feature The deep semantic features are fused with the original deep semantic features to obtain the residual-enhanced deep semantic features, thus completing the multi-scale contextual residual enhancement.

5. The intelligent underwater crack detection and size quantification model as described in claim 1, characterized in that, The acquisition of multi-level fusion features include: The shallow crack detail features are used as the reference features; The mid-level crack structure features and the residual-enhanced deep semantic features are concatenated with the baseline features through channel mapping and upsampling to unify the channel dimension and spatial scale, thereby obtaining multi-level fused features. .

6. The intelligent underwater crack detection and size quantification model as described in claim 5, characterized in that, The generated crack structure response map and refined crack features include: The multi-level fusion features The input is a crack structure perception refinement module (CSRM), and the structure response head of the CSRM is used to refine the multi-level fused features. Mapping is performed to obtain the crack structure response map; The crack structure response map is combined with the multi-level fusion feature. Element-wise multiplication yields the structural enhancement features. Perform 3×3 thinned convolution and fuse it with the multi-level features. By summing the residuals, the refined crack features are obtained. .

7. The intelligent underwater crack detection and size quantification model as described in claim 1, characterized in that, The crack probability map The expression is: ； in, H and W represent the spatial height and width of the image to be recognized, respectively; This is the Sigmoid activation function.

8. The intelligent underwater crack detection and size quantification model as described in claim 1, characterized in that, The process of obtaining the binary segmentation mask M for the crack includes: Set a threshold T, and use the threshold T to analyze the crack probability map. Binarization is performed to obtain the binary segmentation mask M for the crack, which is expressed as: ； Where P(i,j) represents the probability map of pixel (i,j) in the crack. The probability value above.

9. The intelligent underwater crack detection and size quantification model as described in claim 1, characterized in that, The calculation of the area, length, and width of the binary segmentation mask M for the crack includes: Perform connected component analysis on the binary segmentation mask M of the crack, and count the number of connected regions of the crack. Calculate the pixel area of ​​the crack ; Obtain the actual length y corresponding to a single pixel, and calculate the actual area of ​​the crack. ; The binary segmentation mask M of the crack is skeletonized to obtain the crack centerline S and the number of skeleton pixels, and the crack length is calculated. Obtain the pixel calibration ratio z and calculate the actual length of the crack. ; Perform a distance transform on the binary segmentation mask M of the crack to obtain the distance from each pixel within the crack region to the nearest background pixel. Calculate the average width of local cracks and the maximum width of local cracks Based on the pixel calibration ratio z, the actual average width of the local crack is calculated. and the actual maximum width of local cracks .

10. The intelligent underwater crack detection and size quantification model as described in claim 9, characterized in that, The generated visualization images for underwater crack identification and size quantification include: The actual area of ​​the crack Actual length of the crack The actual average width of local cracks and the actual maximum width of local cracks The data is statistically analyzed to generate a comprehensive dimension report. The dimension data in the comprehensive dimension report is then overlaid onto the original underwater dam surface image to generate a visual image for underwater crack identification and dimension quantification.

Images