Computer vision-based method for detecting seawater erosion depth of soft soil foundation solidification layer

Abstract

The present application relates to the technical field of material detection, and discloses a method for detecting seawater erosion depth of a soft soil foundation solidification layer based on computer vision, which comprises the following steps: obtaining a multispectral polarization image sequence of the surface of the soft soil foundation solidification layer through underwater polarization imaging to obtain original data containing polarization degree and spectral characteristics; eliminating seawater flow blur through spatiotemporal filtering of motion vector fields between adjacent frames to generate a stable enhanced image; extracting micro textures and macro profiles based on the image using a multiscale morphological gradient, and constructing an edge confidence map with adaptive interface blur through feature fusion; combining the optical absorption characteristics of the solidification layer material to establish a depth regression model and calculate a seawater erosion depth distribution map; and comparing the seawater erosion depth distribution map with historical erosion data in time sequence, identifying an erosion acceleration area using an outlier detection algorithm, and outputting a comprehensive detection report containing spatial distribution and time evolution trend.

Description

Technical Field

[0001] This invention relates to the field of materials testing technology, and in particular to a method for detecting the depth of seawater erosion in the solidified layer of soft soil foundation based on computer vision. Background Technology

[0002] Existing detection methods cannot acquire multi-dimensional original images using multispectral polarization imaging, lack spatiotemporal joint denoising methods based on motion vectors, are difficult to extract and adaptively fuse texture and contour features through multi-scale layering, and cannot rely on material optical parameters to build regression models to achieve quantitative measurement of global erosion.

[0003] For example, conventional underwater detection only uses single-band imaging to acquire images. The dynamic blurring caused by seawater flow cannot be effectively eliminated. Microscopic erosion patterns and macroscopic boundaries are mixed with a large number of false edges. Relying solely on manual fixed-point sampling to calculate the erosion depth cannot complete the full-area spatial depth mapping, nor can it combine historical data to locate local accelerated erosion blocks.

[0004] Therefore, existing technologies suffer from poor image anti-interference capabilities, one-sided feature extraction, and detection mainly based on single-point sampling, lacking the ability to perform spatial global quantization and temporal evolution analysis. This invention adopts a complete set of differentiated technical paths, including polarization multi-source imaging, spatiotemporal filtering and noise reduction, multi-scale adaptive feature fusion, optical correlation regression modeling, and temporal anomaly discrimination, to overcome the above-mentioned defects. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the purpose of this invention is to propose a computer vision-based method for detecting the depth of seawater erosion in the solidified layer of soft soil foundations. This method aims to solve the technical problems of poor image anti-interference capability, one-sided feature extraction, detection mainly based on single-point sampling, and lack of spatial global quantization and temporal evolution analysis capabilities in existing technologies.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a method for detecting the depth of seawater erosion in the solidified layer of soft soil foundation based on computer vision. The computer vision-based method for detecting the depth of seawater erosion in the solidified layer of soft soil foundation includes: S1. Acquire a multispectral polarization image sequence of the solidified layer surface of soft soil foundation using an underwater polarization imaging device to obtain raw image data containing polarization degree information and spectral features; S2. Perform spatiotemporal filtering on the original image data, and use the motion vector field between adjacent frames to eliminate the dynamic blur caused by seawater flow to generate an interface-enhanced image. S3. Based on the interface enhancement image, the micro-texture features and macro-contour features of the solidified layer surface are extracted using a multi-scale morphological gradient operator, and an interface ambiguity adaptive edge confidence map is constructed using a feature fusion algorithm. S4. Based on the edge confidence map and combined with the optical absorption characteristic parameters of the solidified layer material, construct a depth regression model to calculate the seawater erosion depth distribution map of the solidified layer of soft soil foundation. S5. Compare and analyze the seawater erosion depth distribution map with historical erosion data over time, identify areas of accelerated erosion through an outlier detection algorithm, and output a comprehensive detection report that includes spatial distribution and temporal evolution trends.

[0007] Preferably, the step of acquiring a multispectral polarization image sequence of the solidified layer surface of the soft soil foundation using an underwater polarization imaging device to obtain raw image data containing polarization degree information and spectral features includes: The underwater polarization imaging device is placed in the water above the solidified layer of the soft soil foundation, with the lens of the underwater polarization imaging device facing the surface of the solidified layer. The underwater polarization imaging device is activated, and the polarizer group inside the device switches multiple polarization angles in sequence. Multispectral images are acquired at each polarization angle, and a set of multispectral images associated with polarization angles is generated. The underwater polarization imaging device is controlled to move along the surface of the solidified layer. After each acquisition position is reached, the polarizer group switching and multispectral image acquisition are repeated to obtain a multispectral image set corresponding to multiple acquisition positions. The multispectral image set is correlated according to polarization angle and spectral channels to generate raw image data containing polarization degree information and spectral features at each location.

[0008] Preferably, the original image data undergoes spatiotemporal filtering to eliminate dynamic blur caused by seawater flow using motion vector fields between adjacent frames, generating an interface-enhanced image, including: Block matching is performed on multiple consecutive frames of images in the original image data to obtain the motion vector field between adjacent frames. The motion vector field represents the displacement direction and displacement amplitude of local areas of the image caused by seawater flow. Based on the motion vector field, reverse motion compensation is performed on the current frame image, and corresponding pixels in adjacent frames are aligned to the same spatial reference system to obtain the motion-compensated aligned frame sequence. The pixels at the same spatial location in the aligned frame sequence are subjected to median filtering in the time domain to filter out temporal grayscale anomalies caused by seawater flow, thereby generating a temporally smooth intermediate frame image. Spatial domain bilateral filtering is performed on the intermediate frame image to preserve the edge details of the solidified layer interface and eliminate residual spatial noise, resulting in an interface-enhanced image.

[0009] Preferably, performing temporal median filtering on pixels at the same spatial location in the aligned frame sequence to filter out temporal grayscale anomalies caused by seawater flow, and generating a temporally smooth intermediate frame image, includes: Extract the pixel grayscale values ​​of multiple consecutive frames of images at the same spatial coordinates from the aligned frame sequence to form a temporal grayscale sequence at the corresponding spatial location. All gray values ​​in the time-domain gray-scale sequence are sorted in ascending order according to their numerical values ​​to obtain the sorted gray-scale sequence. Select the gray value located in the middle position from the sorted gray value sequence as the output gray value at the corresponding spatial position to replace the abnormal deviation value in the original temporal gray value sequence; Traverse all spatial positions in the aligned frame sequence, recombine the output grayscale values ​​obtained at each position into a complete image frame, and output a temporally smooth intermediate frame image.

[0010] Preferably, the step of extracting microscopic texture features and macroscopic contour features of the solidified layer surface using a multi-scale morphological gradient operator based on the interface enhancement image, and constructing an interface ambiguity adaptive edge confidence map through a feature fusion algorithm, includes: Morphological dilation and erosion operations are performed on the interface enhancement image by applying the first-scale structuring element and the second-scale structuring element respectively to obtain the first-scale gradient response map and the second-scale gradient response map. The first-scale structuring element is used to capture the micro-texture details of the surface of the cured layer, and the second-scale structuring element is used to capture the macro-contour boundary of the surface of the cured layer. The first-scale gradient response map is subjected to local contrast normalization to enhance the response value of low-contrast regions in the micro-texture features, thereby generating an enhanced micro-texture feature map. Non-maximum suppression processing is applied to the second-scale gradient response map to remove pseudo-edge responses in the macroscopic contour boundary and generate a refined macroscopic contour feature map. The enhanced micro-texture feature map and the refined macro-contour feature map are fused pixel-by-pixel with the local gray-level variance of the interface enhancement image as the fusion weight adjustment factor, and an edge confidence map with adaptive interface ambiguity is output.

[0011] Preferably, the step of performing pixel-by-pixel weighted fusion of the enhanced micro-texture feature map and the refined macro-contour feature map, and using the local gray-level variance of the interface enhancement image as the fusion weight adjustment factor to output an edge confidence map with adaptive interface ambiguity, includes: The interface enhancement image is divided into local windows, and the gray-level variance of pixels in each window is obtained to form a local gray-level variance map that reflects the degree of interface blur. Based on the local grayscale variance map, a fusion weight of the micro-texture feature map and the macro-contour feature map is assigned to each pixel location, wherein the region with larger variance focuses on micro-texture features and the region with smaller variance focuses on macro-contour features, thus obtaining a weight allocation map. The enhanced micro-texture feature map and the refined macro-contour feature map are fused pixel-by-pixel using the weight allocation map, and the fusion result is normalized and mapped to output an edge confidence map that adapts to the interface ambiguity.

[0012] Preferably, the step of constructing an adaptive edge confidence map for interface ambiguity using a feature fusion algorithm includes: Local structural tensor analysis is performed on the interface enhancement image to obtain the ratio of the structural response intensity of the principal direction to the orthogonal direction at the pixel location, which is used as the anisotropy factor at the corresponding location. Based on the combination relationship between the local gray-level variance of each pixel location and the anisotropic factor, a fusion weight between micro-texture features and macro-contour features is generated. Based on the fusion weights, the response values ​​at corresponding positions in the micro-texture feature map and the macro-contour feature map are adaptively weighted and combined to obtain the initial fusion confidence. The initial fusion confidence is subjected to global consistency enhancement processing, which further strengthens the response values ​​in high confidence regions and further suppresses the response values ​​in low confidence regions, outputting an edge confidence map with adaptive interface ambiguity.

[0013] Preferably, the step of constructing a depth regression model based on the edge confidence map and combining it with the optical absorption characteristic parameters of the solidified layer material to calculate the seawater erosion depth distribution map of the soft soil foundation solidified layer includes: Traverse each pixel position in the edge confidence map, read the corresponding edge confidence value, and associate the edge confidence value with the optical absorption characteristic parameters of the pre-calibrated cured layer material to obtain the optical attenuation coefficient of each pixel position; Based on the spatial distribution of the edge confidence map, multiple local image blocks are divided, and the optical attenuation coefficients of all pixels in each local image block are statistically aggregated to generate a local region attenuation feature vector. The local region attenuation feature vector is combined with the edge confidence distribution of the corresponding local image patch to construct the input feature set of the deep regression model with local image patch as the basic unit; The input feature set is input into a pre-established depth regression model block by block. The depth regression model outputs the seawater erosion depth value corresponding to each local image block. The erosion depth values ​​of all local image blocks are stitched together according to their spatial positions in the edge confidence map to generate a seawater erosion depth distribution map of the soft soil foundation solidification layer.

[0014] Preferably, each pixel position in the edge confidence map is traversed, the corresponding edge confidence value is read, and the edge confidence value is correlated and mapped with the optical absorption characteristic parameters of the pre-calibrated cured layer material to obtain the optical attenuation coefficient of each pixel position, including: Traverse each pixel position in the edge confidence map, read the edge confidence value corresponding to the pixel position, and obtain the edge confidence distribution of the entire image; Each edge confidence value in the full image edge confidence distribution is associated and mapped with the optical absorption characteristic parameters of the pre-calibrated cured layer material to obtain the optical attenuation coefficient map at each pixel location; Based on the optical attenuation coefficient map, the edge confidence map is transformed into depth values, mapping the optical attenuation coefficient of each pixel location to the corresponding erosion depth value, and outputting a seawater erosion depth distribution map of the soft soil foundation solidification layer.

[0015] Preferably, the step of performing a time-series comparative analysis of the seawater erosion depth distribution map with historical erosion data, identifying erosion acceleration areas through anomaly detection algorithms, and outputting a comprehensive detection report containing spatial distribution and temporal evolution trends includes: The current seawater erosion depth distribution map is compared with the erosion depth maps of multiple historical times at the corresponding spatial locations in the historical erosion data by performing pixel-by-pixel difference calculations to obtain the time series sequence of depth changes at each spatial location. Sliding window statistical analysis is performed on the depth change time series to calculate the mean and variance of depth change at each spatial location within a preset time window, and a change baseline parameter map is obtained. The depth change value at the current moment is compared with the corresponding position parameter in the change baseline parameter map, and the spatial position where the depth change value exceeds the sum of the change mean and the multiple variance is marked as the candidate region for accelerated erosion. Connectivity analysis is performed on the candidate regions of accelerated erosion to remove isolated noise regions with an area smaller than a preset threshold. The spatial range change trajectory of each connected region is recorded in chronological order, and a comprehensive detection report containing the spatial distribution map of the accelerated erosion region and the temporal evolution trend curve is output.

[0016] The beneficial effects of this invention are as follows: This invention relies on underwater multispectral polarization imaging to complete the raw data acquisition, eliminates image distortion interference caused by water flow by using motion vector field combined with spatiotemporal joint filtering, and extracts multi-level features by relying on multi-scale morphological operation and adaptive weighted fusion strategy, effectively improving the accuracy of solidified layer interface feature extraction and image signal-to-noise ratio, and providing a reliable image data source for erosion depth calculation.

[0017] This invention combines material optical parameters to construct a block-based depth regression model to achieve full-domain erosion depth inversion. Then, relying on temporal difference and sliding statistics, and connected domain screening, it completes the identification of erosion anomaly areas and outputs complete detection results with spatial and temporal information, significantly improving the quantitative detection accuracy and data analysis completeness of seawater erosion depth in soft soil foundations. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a computer vision-based method for detecting the depth of seawater erosion in a solidified layer of soft soil foundation, according to an embodiment of the present invention. Figure 2 A flowchart of time-series erosion anomaly assessment for a computer vision-based method for detecting seawater erosion depth in a solidified layer of soft soil foundation, as provided in an embodiment of the present invention. Figure 3 Microscopic texture feature map of a computer vision-based method for detecting the depth of seawater erosion in a solidified layer of soft soil foundation, provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of the method for detecting the depth of seawater erosion in the solidified layer of soft soil foundation based on computer vision according to the present invention. The first embodiment of the method for detecting the depth of seawater erosion in the solidified layer of soft soil foundation based on computer vision according to the present invention is presented.

[0022] In the first embodiment, the computer vision-based method for detecting the depth of seawater erosion in the solidified layer of soft soil foundation includes: S1. Acquire a multispectral polarization image sequence of the solidified layer surface of soft soil foundation using an underwater polarization imaging device to obtain raw image data containing polarization degree information and spectral features; In this embodiment of the invention, the step of acquiring a multispectral polarization image sequence of the surface of the solidified layer of soft soil foundation using an underwater polarization imaging device to obtain raw image data containing polarization degree information and spectral features includes: The underwater polarization imaging device is placed in the water above the solidified layer of the soft soil foundation, with the lens of the underwater polarization imaging device facing the surface of the solidified layer. The underwater polarization imaging device is activated, and the polarizer group inside the device switches multiple polarization angles in sequence. Multispectral images are acquired at each polarization angle, and a set of multispectral images associated with polarization angles is generated. The underwater polarization imaging device is controlled to move along the surface of the solidified layer. After each acquisition position is reached, the polarizer group switching and multispectral image acquisition are repeated to obtain a multispectral image set corresponding to multiple acquisition positions. The multispectral image set is correlated according to polarization angle and spectral channels to generate raw image data containing polarization degree information and spectral features at each location.

[0023] Specifically, using a matching underwater support frame, the underwater polarization imaging device, having completed the waterproofing test, is placed inside the water above the solidified layer of the soft soil foundation. The height and horizontal orientation of the support frame are fine-tuned until the device's lens is vertically aligned with the surface of the solidified layer, after which the support is locked. The underwater polarization imaging device uses an industrial-grade underwater polarization camera with a resolution of 1920×1080 pixels. It is equipped with an adjustable LCD polarizer group, supporting automatic switching between four polarization angles: 0°, 45°, 90°, and 135°, with a switching time ≤50ms. The multispectral imaging module includes four narrowband spectral channels: 450nm, 532nm, 620nm, and 700nm, each with a bandwidth of ±10nm. The device has a built-in LED ring light source with a color temperature of 5000K and adjustable light intensity; in this embodiment, it is set to 600 lumens. The device moves along a translational track via a stepper motor, with a step size of 0.5 meters and a positioning accuracy of ±1mm. The device was successfully placed at a depth of 1.2 meters in the example site.

[0024] The control equipment for the polarization angle storage repository, which was constructed in advance, was calibrated and formed in a test water environment with the same water quality as the construction area. Four fixed acquisition angles of 0 degrees, 45 degrees, 90 degrees and 135 degrees were selected and determined. The equipment switched polarizer groups according to the storage order in the repository. Under a single polarization angle, the multispectral photosensitive element of the equipment acquired full-band images and bound corresponding parameter tags. After all angles were acquired, a set of multispectral images associated with polarization angles was obtained.

[0025] The underwater polarization imaging equipment is moved laterally along the surface of the solidified layer by a matching translational track. The example single movement distance is 0.7 meters. After the equipment arrives at the designated acquisition point, the track structure is locked, and the polarizer group switching and multispectral image acquisition process is repeated. Each point generates its own corresponding multispectral image set and stores it in the equipment's storage unit.

[0026] Data association processing was performed on the entire multispectral image set. The light intensity values ​​corresponding to the four types of polarization angles under the same spectral channel of the same pixel were extracted in turn. The four basic Stokes parameters were calculated step by step based on the light intensity values. Then, the polarization degree of a single pixel was calculated based on the basic parameters. The spectral data of each point was matched and integrated with the corresponding polarization degree one by one. Finally, the original image data containing the polarization degree information and spectral features of each location was generated.

[0027] This step acquires images by using multiple polarization angles and multiple spectral points to compensate for the lack of information in conventional single-band underwater imaging. It obtains multi-dimensional raw data, providing a complete data source for subsequent spatiotemporal filtering and deblurring, multi-scale feature extraction, and depth regression modeling. This solidifies the data foundation for quantitative detection throughout the entire process and improves the overall accuracy of erosion detection.

[0028] S2. Perform spatiotemporal filtering on the original image data, and use the motion vector field between adjacent frames to eliminate the dynamic blur caused by seawater flow to generate an interface-enhanced image. In this embodiment of the invention, the original image data is subjected to spatiotemporal filtering processing, and the dynamic blur caused by seawater flow is eliminated by utilizing the motion vector field between adjacent frames to generate an interface-enhanced image, including: Block matching is performed on multiple consecutive frames of images in the original image data to obtain the motion vector field between adjacent frames. The motion vector field represents the displacement direction and displacement amplitude of local areas of the image caused by seawater flow. Based on the motion vector field, reverse motion compensation is performed on the current frame image, and corresponding pixels in adjacent frames are aligned to the same spatial reference system to obtain the motion-compensated aligned frame sequence. The pixels at the same spatial location in the aligned frame sequence are subjected to median filtering in the time domain to filter out temporal grayscale anomalies caused by seawater flow, thereby generating a temporally smooth intermediate frame image. Spatial domain bilateral filtering is performed on the intermediate frame image to preserve the edge details of the solidified layer interface and eliminate residual spatial noise, resulting in an interface-enhanced image.

[0029] Performing temporal median filtering on pixels at the same spatial location in the aligned frame sequence to filter out temporal grayscale anomalies caused by seawater flow, and generating a temporally smoothed intermediate frame image, includes: Extract the pixel grayscale values ​​of multiple consecutive frames of images at the same spatial coordinates from the aligned frame sequence to form a temporal grayscale sequence at the corresponding spatial location. All gray values ​​in the time-domain gray-scale sequence are sorted in ascending order according to their numerical values ​​to obtain the sorted gray-scale sequence. Select the gray value located in the middle position from the sorted gray value sequence as the output gray value at the corresponding spatial position to replace the abnormal deviation value in the original temporal gray value sequence; Traverse all spatial positions in the aligned frame sequence, recombine the output grayscale values ​​obtained at each position into a complete image frame, and output a temporally smooth intermediate frame image.

[0030] Specifically, a fixed-size 8-pixel by 8-pixel full-search block matching algorithm is used to split the continuous multi-frame images in the original image data. Each sub-block of the reference frame is traversed in the entire candidate frame to search for matching blocks. The optimal matching position is determined by using the minimum pixel error as the screening criterion. The local pixel offset parameters are statistically calculated using the difference in the coordinates of the matching blocks in the previous and next frames. All sub-block offset data in the whole image are integrated to generate a motion vector field that represents the direction and amplitude of local pixel displacement caused by seawater flow.

[0031] The offset parameters of each region stored in the motion vector field are retrieved, and the spatial coordinates of each pixel in the adjacent frame are translated in reverse with the current frame coordinates as the reference. This makes the pixels of the corresponding solidified layer interface in different frames uniformly mapped to the same spatial reference system. After the coordinates of all pixels in the whole image are corrected, the motion-compensated aligned frame sequence is generated.

[0032] By locking the same spatial coordinate positions within the alignment frame sequence one by one, extracting all gray values ​​corresponding to the coordinates of five consecutive frames to form a temporal gray value sequence, arranging all gray values ​​in a single temporal gray value sequence in ascending order from smallest to largest to obtain a sorted gray value sequence, selecting the middle value of the sorted sequence as the final output gray value of the coordinate to eliminate temporal gray value anomalies caused by seawater disturbance, traversing all coordinates of the image and filling the corresponding output gray values, and then combining them to form a smooth intermediate frame image in the temporal domain.

[0033] The intermediate frame image is processed pixel by pixel using a spatial bilateral filtering algorithm. The algorithm simultaneously calculates the pixel spatial distance weight and gray level difference weight. The neighboring pixels that are closer to the center pixel and have smaller gray level differences have a higher weight ratio. The original pixel values ​​are retained in the gray level change areas at the edge of the solidified layer interface. Scattered noise in the image is smoothed by weighted averaging. After the full image traversal operation is completed, the interface enhancement image is obtained.

[0034] Within a fixed alignment frame sequence, any set of identical pixel spatial coordinates is used. Five consecutive aligned images are selected as a fixed sampling frame number. The pixel grayscale value corresponding to the spatial coordinate in each frame image is accurately read. All the collected grayscale values ​​are collected and arranged in an orderly manner to form a temporal grayscale sequence at the corresponding spatial location. For example, if the grayscale values ​​of a certain coordinate are 128, 196, 130, 129, and 200 in five consecutive frames, the complete collection forms a temporal grayscale sequence specific to that point.

[0035] A uniform ascending sort operation is performed on all pixel grayscale values ​​within the generated temporal grayscale sequence. All data within the sequence are rearranged one by one according to the rule of numerical values ​​from smallest to largest. After the disordered original grayscale data is sorted, an ordered sorted grayscale sequence is generated. The sorted grayscale sequence of the aforementioned example sequence is 128, 129, 130, 196, and 200.

[0036] The sorted grayscale sequence is processed using the median filtering rule of the median value of the odd sequence. When the total number of frames is five, the grayscale value of the third median position in the sorted sequence is fixed and selected. This median grayscale value is set as the final output grayscale value of the pixel at the current spatial position. This stable normal value is used to replace the abnormal deviation values ​​that are too high or too low caused by seawater flow disturbance in the original temporal grayscale sequence. In the example above, the median value of 130 is selected as the output grayscale value of this point, and the two abnormal deviation values ​​of 196 and 200 are removed.

[0037] The algorithm iterates through all pixel spatial positions covered by the aligned frame sequence image row by row and column by column. For each spatial coordinate point, it completes the operation of constructing a temporal grayscale sequence, sorting grayscale, taking the median value, and replacing outliers. The standard output grayscale values ​​corresponding to all spatial positions are then reassembled into a complete single-frame image according to the pixel arrangement structure of the original image. Finally, it outputs a temporally smooth intermediate frame image with uniform image quality and eliminating temporal disturbances from seawater.

[0038] This step effectively eliminates image blurring and grayscale distortion caused by seawater disturbance by combining motion vector compensation with temporal and spatial hierarchical noise reduction, thus optimizing image quality. It constructs a high-quality benchmark image for subsequent multi-scale feature extraction, ensuring the accuracy of the edge confidence map and thereby improving the accuracy of depth regression modeling and global erosion measurement.

[0039] S3. Based on the interface enhancement image, the micro-texture features and macro-contour features of the solidified layer surface are extracted using a multi-scale morphological gradient operator, and an interface ambiguity adaptive edge confidence map is constructed using a feature fusion algorithm. In this embodiment of the invention, the step of extracting micro-texture features and macro-contour features of the solidified layer surface based on the interface enhancement image using a multi-scale morphological gradient operator, and constructing an interface ambiguity adaptive edge confidence map through a feature fusion algorithm includes: Morphological dilation and erosion operations are performed on the interface enhancement image by applying the first-scale structuring element and the second-scale structuring element respectively to obtain the first-scale gradient response map and the second-scale gradient response map. The first-scale structuring element is used to capture the micro-texture details of the surface of the cured layer, and the second-scale structuring element is used to capture the macro-contour boundary of the surface of the cured layer. The first-scale gradient response map is subjected to local contrast normalization to enhance the response value of low-contrast regions in the micro-texture features, thereby generating an enhanced micro-texture feature map. Non-maximum suppression processing is applied to the second-scale gradient response map to remove pseudo-edge responses in the macroscopic contour boundary and generate a refined macroscopic contour feature map. The enhanced micro-texture feature map and the refined macro-contour feature map are fused pixel-by-pixel with the local gray-level variance of the interface enhancement image as the fusion weight adjustment factor, and an edge confidence map with adaptive interface ambiguity is output.

[0040] The step of performing pixel-by-pixel weighted fusion of the enhanced micro-texture feature map and the refined macro-contour feature map, and using the local gray-level variance of the interface enhancement image as the fusion weight adjustment factor, to output an edge confidence map with adaptive interface ambiguity includes: The interface enhancement image is divided into local windows, and the gray-level variance of pixels in each window is obtained to form a local gray-level variance map that reflects the degree of interface blur. Based on the local grayscale variance map, a fusion weight of the micro-texture feature map and the macro-contour feature map is assigned to each pixel location, wherein the region with larger variance focuses on micro-texture features and the region with smaller variance focuses on macro-contour features, thus obtaining a weight allocation map. The enhanced micro-texture feature map and the refined macro-contour feature map are fused pixel-by-pixel using the weight allocation map, and the fusion result is normalized and mapped to output an edge confidence map that adapts to the interface ambiguity.

[0041] The step of constructing an adaptive edge confidence map for interface ambiguity using a feature fusion algorithm includes: Local structural tensor analysis is performed on the interface enhancement image to obtain the ratio of the structural response intensity of the principal direction to the orthogonal direction at the pixel location, which is used as the anisotropy factor at the corresponding location. Based on the combination relationship between the local gray-level variance of each pixel location and the anisotropic factor, a fusion weight between micro-texture features and macro-contour features is generated. Based on the fusion weights, the response values ​​at corresponding positions in the micro-texture feature map and the macro-contour feature map are adaptively weighted and combined to obtain the initial fusion confidence. The initial fusion confidence is subjected to global consistency enhancement processing, which further strengthens the response values ​​in high confidence regions and further suppresses the response values ​​in low confidence regions, outputting an edge confidence map with adaptive interface ambiguity.

[0042] Specifically, two sets of fixed-size circular morphological structural elements are pre-constructed as computational benchmarks. The first-scale structural element is set to a 3×3 pixel specification, which is specifically used to capture micro-texture details such as minor bumps and cracks on the surface of the cured layer. The second-scale structural element is set to a 7×7 pixel specification, which is specifically used to capture macro-contour boundaries such as the overall boundary and large-scale contour of the cured layer. Morphological dilation and erosion operations are performed sequentially on the interface enhancement image using the two sets of structural elements. At a single scale, the gradient response image at the corresponding scale is obtained by subtracting the result of the erosion operation from the result of the dilation operation. Finally, a first-scale gradient response map adapted to micro-features and a second-scale gradient response map adapted to macro-features are generated respectively.

[0043] A local contrast normalization algorithm with a fixed window size of 5×5 pixels is used to perform global pixel-by-pixel processing on the first-scale gradient response map. This algorithm extracts the mean and standard deviation of the gradient response of all pixels within the window, centered on a single target pixel. The original gradient response value of the target pixel is subtracted from the calculated mean of the window, and then divided by the standard deviation of the window to obtain a new normalized response value. This operation unifies the dynamic range of the global gradient response of the image, specifically enhances the gradient response value of low-contrast areas in the microtexture of the solidified layer, retains effective microtexture information throughout the process, and suppresses uniform background noise. After processing, an enhanced microtexture feature map is generated.

[0044] The second-scale gradient response map is processed by the eight-neighborhood non-maximum suppression algorithm. Taking each gradient response pixel in the image as the center, the gradient value of the pixel is compared with the gradient values ​​of its eight neighboring pixels above, below, left, right, and four diagonal pixels. Only the pixel with the largest gradient value in the current local area is retained as the true contour edge point. False edges and redundant edge response points with gradient values ​​smaller than neighboring pixels are directly removed, realizing single-pixel thinning of the macro contour edge and completely removing false contour signals caused by residual seawater noise. After processing, a thinned macro contour feature map is generated.

[0045] The 3×3 pixel local gray-level variance of the interface enhancement image is used as the dynamic fusion weight adjustment factor. The local gray-level variance is calculated by averaging the squared differences between the gray levels of all pixels in the target pixel's neighborhood and the mean gray level of the neighborhood. The weight allocation uses the following formula: Let the local gray-level variance of pixel position (x, y) be... The global maximum variance is The weights of the micro-texture feature maps Weights of macroscopic contour feature maps A higher variance value indicates a higher degree of blur in the corresponding image area. Increased variance emphasizes microscopic texture features; smaller variance indicates a clearer interface. Increase the emphasis on macroscopic contour features. Perform a pixel-by-pixel weighted superposition operation on the enhanced microtexture feature map and the refined macroscopic contour feature map according to this weighting rule: The fusion result is linearly normalized and mapped to 0~255. After completing the global pixel feature fusion calculation, an edge confidence map is generated that adapts to different degrees of blur in different regions and adaptively adjusts the interface blur.

[0046] The interface enhancement image is divided into global non-overlapping traversal using a fixed 3×3 pixel sliding window. For each sliding window, the average gray value of all pixels within the window is calculated first. Then, the square of the difference between each pixel's gray value and the average value is calculated. The average of all squared values ​​is then used to obtain the local gray variance value corresponding to the current window. The gray variance values ​​calculated for all windows in the entire image are mapped and filled according to their corresponding pixel spatial locations to generate a complete local gray variance map that reflects the degree of interface blur in each region of the image. The larger the variance value, the more severe the residual blur of seawater disturbance in the corresponding region; the smaller the variance value, the clearer the interface image in the corresponding region.

[0047] Based on the variance values ​​of each pixel in the local grayscale variance map, dual feature weight allocation is completed. A fixed variance threshold is preset as the weight boundary standard. For blurry areas in the overall image where the pixel variance value is greater than the preset threshold, the weight of the enhanced micro-texture feature map is increased and the weight of the refined macro-contour feature map is decreased to focus on the details of micro-texture features. For clear areas where the pixel variance value is less than the preset threshold, the weight of the refined macro-contour feature map is increased and the weight of the enhanced micro-texture feature map is decreased to focus on the boundaries of macro-contour features. After all pixels have completed the bidirectional weight assignment, a weight allocation map that completely covers the entire image is generated.

[0048] Read the two sets of weight coefficients corresponding to each pixel in the weight allocation map, multiply the pixel gray value of the enhanced micro-texture feature map with the micro-weight coefficient, and multiply the pixel gray value of the refined macro-contour feature map with the macro-weight coefficient. Add the two sets of multiplied values ​​to obtain the single-pixel fusion value. Traverse all pixels of the image to complete the global pixel-by-pixel weighted fusion operation. Map all pixel fusion values ​​to the standard image gray range of 0 to 255 to complete the normalization process, eliminate the amplitude deviation of the fusion values, and finally output an edge confidence map with adaptive interface ambiguity capability.

[0049] A fixed 3×3 pixel window is used to perform local structural tensor analysis on the interface enhancement image pixel by pixel. The horizontal and vertical gray-level gradients within the window are calculated and a second-order structural tensor matrix is ​​constructed. The matrix is ​​eigenvalued to obtain two sets of structural response intensities corresponding to the principal direction and the orthogonal direction. The structural response intensity value in the principal direction is divided by the structural response intensity value in the orthogonal direction. The result is the anisotropy factor at the current pixel position. The anisotropy factor value is larger in flat contour areas and closer to one in messy texture areas.

[0050] The local gray-level variance of the interface enhancement image is calculated pixel by pixel using a 3×3 pixel window gray-level statistical method. This local gray-level variance is used to accurately characterize the degree of local blurring of the interface. The larger the variance, the more severe the blurring caused by seawater disturbance. The anisotropy factor is used to characterize the determinacy of local edge direction. The larger the factor, the more regular and reliable the edge direction. The local gray-level variance and anisotropy factor of the same pixel are linearly coupled and calculated. Based on the combination relationship between the degree of blurring and the determinacy of the edge, the weights of micro-texture features and macro-contour features are assigned pixel by pixel to generate the fusion weights corresponding to the global pixels.

[0051] The response values ​​at the same pixel position in the micro-texture feature map and the macro-contour feature map are multiplied by the corresponding fusion weights and then summed to complete the single-pixel adaptive weighted combination operation. The global feature fusion calculation is completed by traversing all pixels of the image to obtain the complete initial fusion confidence.

[0052] Global consistency enhancement processing is performed on the initial fusion confidence. The average confidence of the whole image is used as a fixed boundary threshold. The response values ​​of high confidence regions above the mean are amplified to enhance the real edge and texture features, while the response values ​​of low confidence regions below the mean are compressed to suppress noise and pseudo features. After the global correction is completed, the edge confidence map with adaptive interface ambiguity is output.

[0053] This step extracts both microscopic and macroscopic features in a layered manner, and adaptively adjusts the fusion weights based on grayscale variance and anisotropic factors to eliminate false edges and optimize feature reliability. It accurately generates edge confidence maps, providing reliable feature data for subsequent optical parameter mapping and depth regression modeling, ensuring the accuracy of global erosion depth measurement.

[0054] S4. Based on the edge confidence map and combined with the optical absorption characteristic parameters of the solidified layer material, construct a depth regression model to calculate the seawater erosion depth distribution map of the solidified layer of soft soil foundation. In this embodiment of the invention, the step of constructing a depth regression model based on the edge confidence map and combining it with the optical absorption characteristic parameters of the solidified layer material to calculate the seawater erosion depth distribution map of the soft soil foundation solidified layer includes: Traverse each pixel position in the edge confidence map, read the corresponding edge confidence value, and associate the edge confidence value with the optical absorption characteristic parameters of the pre-calibrated cured layer material to obtain the optical attenuation coefficient of each pixel position; Based on the spatial distribution of the edge confidence map, multiple local image blocks are divided, and the optical attenuation coefficients of all pixels in each local image block are statistically aggregated to generate a local region attenuation feature vector. The local region attenuation feature vector is combined with the edge confidence distribution of the corresponding local image patch to construct the input feature set of the deep regression model with local image patch as the basic unit; The input feature set is input into a pre-established depth regression model block by block. The depth regression model outputs the seawater erosion depth value corresponding to each local image block. The erosion depth values ​​of all local image blocks are stitched together according to their spatial positions in the edge confidence map to generate a seawater erosion depth distribution map of the soft soil foundation solidification layer.

[0055] The edge confidence map is traversed, and the corresponding edge confidence value is read. This edge confidence value is then correlated and mapped with pre-calibrated optical absorption characteristic parameters of the cured layer material to obtain the optical attenuation coefficient for each pixel location, including: Traverse each pixel position in the edge confidence map, read the edge confidence value corresponding to the pixel position, and obtain the edge confidence distribution of the entire image; Each edge confidence value in the full image edge confidence distribution is associated and mapped with the optical absorption characteristic parameters of the pre-calibrated cured layer material to obtain the optical attenuation coefficient map at each pixel location; Based on the optical attenuation coefficient map, the edge confidence map is transformed into depth values, mapping the optical attenuation coefficient of each pixel location to the corresponding erosion depth value, and outputting a seawater erosion depth distribution map of the soft soil foundation solidification layer.

[0056] It is important to note that the erosion of the solidified layer in soft soil foundations by seawater begins at the surface, gradually leading to a loosening of the internal cementitious structure, the expansion of microcracks, and the dissolution of mineral components. This alters both the microscopic geometry and macroscopic integrity of the solidified layer surface. The greater the erosion depth, the more complex the surface microtexture becomes, characterized by more fine edge features, and the more blurred the macroscopic boundary becomes, resulting in a weakened gradient response. These changes significantly affect the absorption and scattering characteristics of the solidified layer material to incident light: a slightly eroded solidified layer has a dense and smooth surface with a low optical absorption coefficient, and light intensity attenuation primarily follows the Lambert-Beer law; a severely eroded solidified layer has a rough and porous surface, resulting in multiple scattering of incident light, an increased effective optical path, and a higher optical attenuation coefficient. Furthermore, under polarization imaging, changes in the surface microstructure of the eroded area cause a decrease in polarization degree and alterations in edge response intensity. Therefore, by extracting edge confidence features related to surface micro-texture and macro-contour in the image, the changes in material optical absorption characteristics caused by erosion can be indirectly characterized, and a quantitative mapping relationship from edge confidence to optical attenuation coefficient to erosion depth can be established, realizing vision-based global erosion depth inversion.

[0057] A calibration relationship for the optical absorption characteristics of the cured layer material was established beforehand. The fitting coefficients used in this embodiment are applicable to cured layers on ordinary silicate cement-based soft soil foundations: a water-cement ratio of 0.5 and a curing agent dosage of 15%. The calibration process involved preparing eight sets of standard cured layer samples with different erosion depths (0mm, 5mm, 10mm, 15mm, 20mm, 30mm, 40mm, 50mm). These samples were collected and edge confidence scores were calculated under the same underwater imaging environment. Simultaneously, a spectrophotometer was used to measure the optical attenuation coefficients corresponding to each depth. Unit: cm -1 The quadratic function relationship is obtained by fitting using the least squares method: Where C is the edge confidence score (normalized value from 0 to 1), and the fitting coefficients are a=0.25, b=0.12, and c=0.03. The edge confidence score is read by traversing the global pixel coordinates of the edge confidence map pixel by pixel. The edge confidence score of each pixel is then substituted into the quadratic function to complete the numerical transformation, accurately obtaining the optical attenuation coefficient corresponding to each pixel position, and generating an optical attenuation coefficient map.

[0058] It should be noted that for other materials, the above calibration steps should be repeated: prepare at least 8 sets of corresponding material standard samples with different erosion depths, measure their edge confidence and optical attenuation coefficients, and use the least squares method to fit the coefficients a, b, and c applicable to the corresponding materials.

[0059] The overall imaging region of the edge confidence map is spatially uniformly divided into square, non-overlapping blocks of fixed size, resulting in several locally image blocks of uniform size (e.g., 32×32 pixels) arranged continuously. Statistical aggregation operations are performed on the optical attenuation coefficients corresponding to all pixels within any locally image block. Specifically, the arithmetic mean, standard deviation, maximum value, minimum value, and median of all optical attenuation coefficients within the block are calculated, representing five statistical features. Then, the mean, variance, and skewness of the edge confidence scores within the same image block are calculated, representing three statistical features. These eight features are arranged and combined in a fixed order to generate a unique local region attenuation feature vector corresponding to each locally image block.

[0060] The deep regression model is a three-layer fully connected neural network with the following structure: 8 nodes in the input layer, 64 nodes in the first hidden layer (ReLU activation function, followed by a Dropout layer with a dropout rate of 0.2%), 32 nodes in the second hidden layer (ReLU activation function), and 1 node in the output layer (linear activation, outputting erosion depth values ​​in mm). The loss function is mean squared error (MSE), the optimizer is Adam, the initial learning rate is 0.001, the batch size is 32, the training epochs are 200, and the early stopping epochs are 20. The model training method involves preparing 50 sets of standard samples of soft soil foundation solidification layers in the laboratory. Each set of samples has a known uniform erosion depth (range 0~50 mm, step size 1 mm). Images are acquired, edge confidence maps are extracted, and local image patches are divided in the same way as in actual measurements to construct the input feature set and corresponding true depth labels. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. Mini-batch gradient descent is used during training. The validation set loss is calculated after each training epoch. If the loss does not decrease for 20 consecutive epochs, training is stopped, and the optimal model weights are saved. After the model is trained, all weights and bias parameters are fixed for actual detection.

[0061] The completed input feature set is input into the depth regression model in groups as individual local image patches. The model outputs the accurate seawater erosion depth value corresponding to each local image patch through a fixed forward propagation operation logic. The seawater erosion depth values ​​corresponding to all local image patches are strictly stitched together according to their original spatial positions in the edge confidence map to form a complete two-dimensional depth distribution image, and finally generate the seawater erosion depth distribution map of the soft soil foundation solidification layer.

[0062] The edge confidence map is traversed row by row and column by column, and the edge confidence value stored in each pixel is extracted. All read data is summarized according to the original spatial arrangement of the pixels, and after regular arrangement, a full-image edge confidence distribution that completely covers the entire image area is formed.

[0063] Multiple standard samples of cured layers with different erosion thicknesses were prepared in the laboratory beforehand. The edge confidence and corresponding optical absorption parameters of each sample were measured in an underwater environment consistent with the field. A fixed mapping reference table was generated as the optical absorption characteristic parameters of the pre-calibrated cured layer material. The values ​​of each pixel in the edge confidence distribution of the whole image were retrieved sequentially to complete the conversion. After all pixels were converted, a complete optical attenuation coefficient map was generated according to the spatial arrangement of the original image.

[0064] The fixed conversion relationship between the pre-calibrated optical attenuation coefficient and erosion depth is retrieved. The coefficient values ​​in the optical attenuation coefficient map are taken pixel by pixel to complete the one-to-one conversion. The coefficients of each point are uniformly converted into the erosion depth values ​​of the corresponding positions. After all pixels are converted, the image data is reorganized according to the spatial position of the original image, and the seawater erosion depth distribution map of the soft soil foundation solidification layer is directly output.

[0065] This step utilizes material optical parameters to map confidence levels to attenuation coefficients, constructs feature input regression models in blocks, and achieves quantitative mapping of the entire erosion depth. It overcomes the limitations of traditional single-point sampling and accurately generates depth distribution maps.

[0066] S5. Compare and analyze the seawater erosion depth distribution map with historical erosion data over time, identify areas of accelerated erosion through an outlier detection algorithm, and output a comprehensive detection report that includes spatial distribution and temporal evolution trends.

[0067] In this embodiment of the invention, the step of performing a time-series comparative analysis of the seawater erosion depth distribution map with historical erosion data, identifying erosion acceleration areas through an outlier detection algorithm, and outputting a comprehensive detection report containing spatial distribution and temporal evolution trends includes: The current seawater erosion depth distribution map is compared with the erosion depth maps of multiple historical times at the corresponding spatial locations in the historical erosion data by performing pixel-by-pixel difference calculations to obtain the time series sequence of depth changes at each spatial location. Sliding window statistical analysis is performed on the depth change time series to calculate the mean and variance of depth change at each spatial location within a preset time window, and a change baseline parameter map is obtained. The depth change value at the current moment is compared with the corresponding position parameter in the change baseline parameter map, and the spatial position where the depth change value exceeds the sum of the change mean and the multiple variance is marked as the candidate region for accelerated erosion. Connectivity analysis is performed on the candidate regions of accelerated erosion to remove isolated noise regions with an area smaller than a preset threshold. The spatial range change trajectory of each connected region is recorded in chronological order, and a comprehensive detection report containing the spatial distribution map of the accelerated erosion region and the temporal evolution trend curve is output.

[0068] It should be noted that the historical erosion data is obtained in the following ways: The first type is accumulated data from regular inspections. For soft soil foundation reinforcement projects such as offshore wind power foundations and cross-sea bridge piers that have been in operation for a long time, steps S1 to S4 are repeated in the same inspection area at fixed time intervals (such as every quarter or every six months) using the same equipment parameters and imaging path to generate seawater erosion depth distribution maps for each time point, which are then stored in a time-series database. In this embodiment, it is preferable to collect data once every two weeks to capture short-term fluctuations in the erosion rate.

[0069] The second category is initial baseline data. For newly constructed or recently solidified soft soil foundations, a high-precision full-area scan is conducted before use (i.e., the theoretical state of zero erosion depth) to serve as a time zero-point reference. If an ideal zero point cannot be obtained, the average of the first three test results is used as the baseline. All historical data are archived in chronological order of collection and accompanied by metadata tags, including environmental parameters such as collection time, water temperature, salinity, and pH value, for use in subsequent time-series analysis to remove outliers.

[0070] The real-time seawater erosion depth distribution map is spatially registered with the pre-archived historical erosion depth maps of the same detection area and the same pixel resolution. This ensures that the spatial coordinates of all time-series images correspond one-to-one. For all overlapping spatial locations, a pixel-by-pixel depth value subtraction operation is performed. The current erosion depth value is subtracted from the corresponding historical erosion depth value. All time-series difference results are summarized pixel by pixel and then arranged to form a unique depth change time-series sequence for each spatial location.

[0071] A fixed-duration sliding window is used to perform traversal statistical analysis on the temporal sequence of depth changes corresponding to each pixel. The sliding window length is fixed at 5 time points, for example, data is collected every two weeks, with a window span of 10 weeks. The depth change values ​​of five consecutive time points are included in a single window. The arithmetic mean of all depth change values ​​within the window is calculated pixel by pixel as the depth change mean. At the same time, the squared average of the deviations of the values ​​within the window from the mean is calculated as the change variance. After all pixels have completed the statistical calculations, a baseline parameter map of the change covering the entire area is generated.

[0072] The mean and variance of depth change for each pixel in the baseline parameter map are retrieved. Anomaly identification is carried out using a fixed multiple variance judgment rule. The judgment threshold is fixed as the mean depth change plus three times the variance. The current depth change value of each spatial location is compared with the judgment threshold one by one. All pixel coordinates with current depth change values ​​greater than the threshold are marked and filtered out. The set of all marked pixels constitutes the candidate region for accelerated erosion in the entire domain.

[0073] An eight-connected-domain analysis method was used to segment and statistically analyze the selected candidate regions for accelerated erosion. A preset effective region area threshold of 50 pixels was established, corresponding to an actual area of ​​approximately 0.01 square meters, determined based on the conversion relationship between image resolution and actual scale. The total pixel area of ​​each independent connected domain was automatically calculated, and isolated noise regions and invalid scattered regions with a total pixel area less than the preset threshold were directly removed, retaining the effective accelerated erosion connected regions that met the area standard. Archived full-time erosion data was retrieved, and the dynamic changes in the spatial range, area, and erosion depth of each effective connected domain were continuously recorded chronologically. The final retained spatial distribution data of accelerated erosion regions and the entire time evolution trajectory data were integrated to generate a comprehensive detection report that simultaneously includes a spatial distribution map of accelerated erosion regions and a time evolution trend curve.

[0074] This step constructs an erosion baseline based on temporal difference and sliding statistics, and accurately locates accelerated erosion regions by combining threshold screening and connected component noise reduction, thus overcoming the shortcomings of traditional single-point detection. It completes the entire detection loop, outputting an integrated spatial and temporal report, improving the overall practicality of the detection and the completeness of data analysis.

[0075] Example 2 In this application, the computer vision-based method for detecting the depth of seawater erosion in the solidified layer of soft soil foundation includes: Images of the solidified layer are obtained by underwater multispectral polarization imaging. Spatiotemporal filtering is used to eliminate interference from seawater flow. Then, an erosion depth distribution map is generated by a depth regression model to complete the non-destructive quantitative detection of the erosion status of the solidified layer.

[0076] Specifically, the device uses a remotely operated vehicle (ROV) as its platform and is equipped with an industrial-grade underwater polarization camera with a resolution of 1920 x 1080 pixels. It features an adjustable LCD polarizer group that supports automatic switching between four polarization angles: 0°, 45°, 90°, and 135°, with a switching time of no more than 50 milliseconds. The multispectral imaging module includes four narrowband spectral channels: 450 nm, 532 nm, 620 nm, and 700 nm, each with a bandwidth of plus or minus 10 nm. The device incorporates a built-in LED ring light source with a color temperature of 5000 Kelvin and an illumination intensity of 600 lumens.

[0077] The ROV was lowered to a depth of approximately 2.0 meters in water around the substation's pile foundation, with its lens facing the surface of the solidified layer. Multiple acquisition points were set along the solidified layer surface in 0.5-meter increments. At each point, four polarization angles were sequentially switched, and four-channel multispectral images were acquired for each point. The exposure time for a single point, single angle, and single channel was 50 milliseconds, and the acquisition time for four channels at a single angle was approximately 0.2 seconds, for a total of approximately 0.8 seconds across the four polarization angles. Moving to the next point took approximately 15 seconds, and complete data acquisition for approximately 50 points across the entire substation area was completed within about 3 hours.

[0078] The polarization information of each pixel is calculated and correlated with the spectral channel data to generate the original image data, which includes the omnidirectional polarization characteristics and multispectral radiation characteristics at each solidified layer location.

[0079] The original image sequence comprises approximately 2000 frames, with about 40% to 60% of these frames affected by motion blur caused by seawater flow, manifesting as ghosting of the solidified layer outline and trailing of micro-textures. Block matching was performed on consecutive frames with an 8x8 pixel search window to obtain the motion vector field between adjacent frames. The amplitude distribution of the motion vectors across the entire image was statistically analyzed; the vector amplitude ranged from approximately 1 to 8 pixels, with random directions. After inverse motion compensation, an aligned frame sequence was obtained. Then, the grayscale values ​​of pixels with the same spatial coordinates were grouped into a temporal sequence, and the median value was used to replace outliers. After temporal median filtering, a spatial domain bilateral filter was used for global smoothing, preserving areas of abrupt grayscale changes, i.e., the edges of the solidified layer interface and cracks, while simultaneously eliminating incompletely removed speckle noise.

[0080] The enhanced underwater image quality index improved from approximately 0.4 in the original image to over 1.2, and the average gradient improved from approximately 2.3 to over 7.5, clearly identifying the surface texture and erosion boundaries of the solidified layer.

[0081] Morphological dilation and erosion were performed on the interface enhancement image using a first-scale structuring element (3x3 pixels) and a second-scale structuring element (7x7 pixels). The small-scale gradient response map preserved micro-texture features such as the edges of dissolution pits, micro-crack details, and particle peeling traces, while the large-scale gradient response map extracted macro-contour boundaries such as the overall thickness transition zone of the solidified layer, the pile foundation interface zone, and the contours of uneven regional settlement. Low-contrast regions in the micro-features were enhanced by local contrast normalization, and false edges in the macro-gradient map were removed by eight-neighborhood non-maximum suppression. Finally, the fusion weights of micro- and macro-features were adaptively adjusted using local gray-level variance: blurred regions with high variance emphasized micro-texture, while clear regions with low variance emphasized macro-contours, generating an edge confidence map.

[0082] For the silicate cement-based curing layer material used in the Rudong project (water-cement ratio of 0.45, curing agent dosage of 12%), optical absorption characteristic parameters were calibrated according to a patented method: Nine sets of standard samples with different erosion depths were prepared, namely 0 mm, 2 mm, 5 mm, 10 mm, 15 mm, 20 mm, 30 mm, 40 mm, and 50 mm, with three parallel samples in each set. Multispectral polarization images were acquired and edge confidence scores were extracted in a simulated underwater environment (temperature 15 degrees Celsius, salinity 0.28%, turbidity 300 NTU). Simultaneously, the optical attenuation coefficient corresponding to each depth was measured using a spectrophotometer. A quadratic function relationship between the edge confidence score and the optical attenuation coefficient was obtained by least squares fitting, with a correlation coefficient reaching 0.97.

[0083] The edge confidence map is divided into multiple 32x32 pixel local image patches. The mean, standard deviation, maximum, minimum, and median of the optical attenuation coefficients within each patch, as well as the mean, variance, and skewness of the edge confidence scores, are calculated to form an eight-dimensional feature vector. These features are then input patch by patch into a pre-trained deep regression model, which is a three-layer fully connected neural network: an input layer (eight nodes), a first hidden layer (sixty-four nodes, activated by a linear rectified function with a dropout rate of 0.2), a second hidden layer (thirty-two nodes, activated by a linear rectified function), and an output layer (single node, outputting the erosion depth value). The model outputs the seawater erosion depth for each local image patch.

[0084] The predicted depth values ​​were stitched together according to spatial location to generate a complete erosion depth distribution map. Different color gradations were used to indicate erosion depth ranges: blue-green areas corresponded to 0-10 mm for slight erosion, yellow-orange areas to 10-25 mm for moderate erosion, and reddish-brown areas to greater than 25 mm for severe erosion. The average absolute error between the predicted depth values ​​and the measured values ​​of the validation group samples was controlled within 2.0 mm, with a maximum error not exceeding 3.5 mm.

[0085] Multiple data points are collected at fixed time intervals. The currently generated erosion depth distribution map is compared pixel-by-pixel with historical depth maps from multiple periods to obtain a temporal sequence of depth changes at each location. A change baseline is established through sliding window statistical analysis, and locations where the current change exceeds the baseline threshold are marked as candidate regions for accelerated erosion. After removing isolated noise points through connected component analysis, effective accelerated erosion regions are obtained, and the causes of accelerated erosion can be confirmed by combining high-resolution imaging.

[0086] like Figure 2 The diagram shows a flowchart of the time-series erosion anomaly assessment method for seawater erosion depth detection of solidified layers in soft soil foundations based on computer vision, provided by an embodiment of the present invention. It demonstrates the complete steps from historical data comparison to the output of the detection report. The method generates time-series change data based on pixel-by-pixel differences in multi-period depth maps, constructs a change baseline through sliding window statistics, filters accelerated erosion candidate areas through variance thresholding, removes small-area noise points through connected component filtering, and finally summarizes the regional spatial distribution and evolution curves to form a comprehensive report, thus completing the intelligent identification of accelerated erosion areas.

[0087] like Figure 3The image shown is a microscopic texture feature map of a computer vision-based method for detecting the depth of seawater erosion in a solidified layer of soft soil foundation, provided by an embodiment of the present invention. The left side of the image shows the most severely eroded area, exhibiting a dense, porous network structure resembling fine veins and a sponge skeleton. This is a complex fractal texture formed by the remaining skeleton after long-term seawater dissolution and stripping away the easily soluble components of the material. The erosion level gradually decreases in the middle section, with only sparse microcracks and scattered dissolution pits, reflecting the transitional characteristics of dissolution from the surface inwards. The right side shows a dense and uniform uneroded substrate, with the three forming a distinct erosion front boundary. The image, relying on the contrast between light and dark areas and the structure, intuitively presents the gradual pattern of seawater erosion and the failure characteristics of the solidified layer material in a seawater environment.

[0088] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0089] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A computer vision-based method for detecting the depth of seawater erosion in the solidified layer of soft soil foundation, characterized in that, The method includes: S1. Acquire a multispectral polarization image sequence of the surface of the solidified layer of soft soil foundation using an underwater polarization imaging device to obtain raw image data containing polarization degree information and spectral features; S2. Perform spatiotemporal filtering on the original image data, and use the motion vector field between adjacent frames to eliminate the dynamic blur caused by seawater flow to generate an interface-enhanced image. S3. Based on the interface enhancement image, a multi-scale morphological gradient operator is used to extract the micro-texture features and macro-contour features of the solidified layer surface. An interface ambiguity adaptive edge confidence map is constructed using a feature fusion algorithm, including: Morphological dilation and erosion operations are performed on the interface enhancement image by applying the first-scale structuring element and the second-scale structuring element respectively to obtain the first-scale gradient response map and the second-scale gradient response map. The first-scale structuring element is used to capture the micro-texture details of the surface of the cured layer, and the second-scale structuring element is used to capture the macro-contour boundary of the surface of the cured layer. The first-scale gradient response map is subjected to local contrast normalization to enhance the response value of low-contrast regions in the micro-texture features, thereby generating an enhanced micro-texture feature map. Non-maximum suppression processing is applied to the second-scale gradient response map to remove pseudo-edge responses in the macroscopic contour boundary and generate a refined macroscopic contour feature map. The enhanced micro-texture feature map and the refined macro-contour feature map are fused pixel by pixel with weighted fusion, and the local gray-level variance of the interface enhancement image is used as the fusion weight adjustment factor to output an edge confidence map with adaptive interface ambiguity. S4. Based on the edge confidence map and combined with the optical absorption characteristic parameters of the solidified layer material, construct a depth regression model to calculate the seawater erosion depth distribution map of the solidified layer in soft soil foundation, including: Traverse each pixel position in the edge confidence map, read the corresponding edge confidence value, and associate the edge confidence value with the optical absorption characteristic parameters of the pre-calibrated cured layer material to obtain the optical attenuation coefficient of each pixel position; Based on the spatial distribution of the edge confidence map, multiple local image blocks are divided, and the optical attenuation coefficients of all pixels in each local image block are statistically aggregated to generate a local region attenuation feature vector. The local region attenuation feature vector is combined with the edge confidence distribution of the corresponding local image patch to construct the input feature set of the deep regression model with local image patches as the basic units; The input feature set is input into a pre-established depth regression model block by block. The depth regression model outputs the seawater erosion depth value corresponding to each local image block. The erosion depth values ​​of all local image blocks are spliced ​​together according to their spatial positions in the edge confidence map to generate a seawater erosion depth distribution map of the soft soil foundation solidification layer. S5. Compare and analyze the seawater erosion depth distribution map with historical erosion data over time, identify areas of accelerated erosion through an outlier detection algorithm, and output a comprehensive detection report that includes spatial distribution and temporal evolution trends.

2. The method for detecting the depth of seawater erosion in the solidified layer of soft soil foundation based on computer vision as described in claim 1, characterized in that, The process involves acquiring a multispectral polarization image sequence of the solidified layer surface of the soft soil foundation using an underwater polarization imaging device, resulting in raw image data containing polarization degree information and spectral features, including: The underwater polarization imaging device is placed in the water above the solidified layer of the soft soil foundation, with the lens of the underwater polarization imaging device facing the surface of the solidified layer. The underwater polarization imaging device is activated, and the polarizer group inside the device switches multiple polarization angles in sequence. Multispectral images are acquired at each polarization angle, and a set of multispectral images associated with polarization angles is generated. The underwater polarization imaging device is controlled to move along the surface of the solidified layer. After each acquisition position is reached, the polarizer group switching and multispectral image acquisition are repeated to obtain a multispectral image set corresponding to multiple acquisition positions. The multispectral image set is correlated according to polarization angle and spectral channels to generate raw image data containing polarization degree information and spectral features at each location.

3. The method for detecting the depth of seawater erosion in the solidified layer of soft soil foundation based on computer vision as described in claim 1, characterized in that, The original image data undergoes spatiotemporal filtering to eliminate dynamic blur caused by seawater flow using motion vector fields between adjacent frames, generating an interface-enhanced image, including: Block matching is performed on multiple consecutive frames of images in the original image data to obtain the motion vector field between adjacent frames. The motion vector field represents the displacement direction and displacement amplitude of local areas of the image caused by seawater flow. Based on the motion vector field, reverse motion compensation is performed on the current frame image, and corresponding pixels in adjacent frames are aligned to the same spatial reference system to obtain the motion-compensated aligned frame sequence. The pixels at the same spatial location in the aligned frame sequence are subjected to temporal median filtering to filter out temporal grayscale anomalies caused by seawater flow, generating a temporally smooth intermediate frame image. Spatial domain bilateral filtering is performed on the intermediate frame image to preserve the edge details of the solidified layer interface and eliminate residual spatial noise, resulting in an interface-enhanced image.

4. The method for detecting the depth of seawater erosion in the solidified layer of soft soil foundation based on computer vision as described in claim 3, characterized in that, Performing temporal median filtering on pixels at the same spatial location in the aligned frame sequence to filter out temporal grayscale anomalies caused by seawater flow, and generating a temporally smoothed intermediate frame image, includes: Extract the pixel grayscale values ​​of multiple consecutive frames of images at the same spatial coordinates from the aligned frame sequence to form a temporal grayscale sequence at the corresponding spatial location. All gray values ​​in the time-domain gray-scale sequence are sorted in ascending order according to their numerical values ​​to obtain the sorted gray-scale sequence. Select the gray value located in the middle position from the sorted gray value sequence as the output gray value at the corresponding spatial position to replace the abnormal deviation value in the original temporal gray value sequence; Traverse all spatial positions in the aligned frame sequence, recombine the output grayscale values ​​obtained at each position into a complete image frame, and output a temporally smooth intermediate frame image.

5. The method for detecting the depth of seawater erosion in the solidified layer of soft soil foundation based on computer vision as described in claim 1, characterized in that, The step of performing pixel-by-pixel weighted fusion of the enhanced micro-texture feature map and the refined macro-contour feature map, and using the local gray-level variance of the interface enhancement image as the fusion weight adjustment factor, to output an edge confidence map with adaptive interface ambiguity includes: The interface enhancement image is divided into local windows, and the gray-level variance of pixels in each window is obtained to form a local gray-level variance map that reflects the degree of interface blur. Based on the local grayscale variance map, a fusion weight of the micro-texture feature map and the macro-contour feature map is assigned to each pixel location, wherein the region with larger variance focuses on micro-texture features and the region with smaller variance focuses on macro-contour features, thus obtaining a weight allocation map. The enhanced micro-texture feature map and the refined macro-contour feature map are fused pixel-by-pixel using the weight allocation map, and the fusion result is normalized and mapped to output an edge confidence map that adapts to the interface ambiguity.

6. The method for detecting the depth of seawater erosion in the solidified layer of soft soil foundation based on computer vision as described in claim 5, characterized in that, The step of constructing an adaptive edge confidence map for interface ambiguity using a feature fusion algorithm includes: Local structural tensor analysis is performed on the interface enhancement image to obtain the ratio of structural response intensity in the principal direction to the orthogonal direction at the pixel location, which is used as the anisotropy factor at the corresponding location. Based on the combination relationship between the local gray-level variance of each pixel location and the anisotropic factor, a fusion weight between micro-texture features and macro-contour features is generated. Based on the fusion weights, the response values ​​at corresponding positions in the micro-texture feature map and the macro-contour feature map are adaptively weighted and combined to obtain the initial fusion confidence. The initial fusion confidence is subjected to global consistency enhancement processing, which further strengthens the response values ​​in high confidence regions and further suppresses the response values ​​in low confidence regions, and outputs an edge confidence map with adaptive interface ambiguity.

7. The method for detecting the depth of seawater erosion in the solidified layer of soft soil foundation based on computer vision as described in claim 1, characterized in that, The edge confidence map is traversed for each pixel position, and the corresponding edge confidence value is read. This edge confidence value is then correlated and mapped with pre-calibrated optical absorption characteristic parameters of the cured layer material to obtain the optical attenuation coefficient for each pixel position, including: Traverse each pixel position in the edge confidence map, read the edge confidence value corresponding to the pixel position, and obtain the edge confidence distribution of the entire image; Each edge confidence value in the full-image edge confidence distribution is correlated and mapped with the optical absorption characteristic parameters of the pre-calibrated cured layer material to obtain the optical attenuation coefficient map at each pixel location; Based on the optical attenuation coefficient map, the edge confidence map is transformed into depth values, mapping the optical attenuation coefficient of each pixel location to the corresponding erosion depth value, and outputting a seawater erosion depth distribution map of the soft soil foundation solidification layer.

8. The method for detecting the depth of seawater erosion in the solidified layer of soft soil foundation based on computer vision as described in claim 1, characterized in that, The process involves comparing the seawater erosion depth distribution map with historical erosion data over time, identifying areas of accelerated erosion using an outlier detection algorithm, and outputting a comprehensive detection report that includes spatial distribution and temporal evolution trends. The current seawater erosion depth distribution map is compared with the erosion depth maps of multiple historical times at the corresponding spatial locations in the historical erosion data by performing pixel-by-pixel difference calculations to obtain the time series sequence of depth changes at each spatial location. Sliding window statistical analysis is performed on the depth change time series to calculate the mean and variance of depth change at each spatial location within a preset time window, and a change baseline parameter map is obtained. The depth change value at the current moment is compared with the corresponding position parameter in the change baseline parameter map, and the spatial position where the depth change value exceeds the sum of the change mean and the multiple variance is marked as the candidate region for accelerated erosion. Connectivity analysis is performed on the candidate regions of accelerated erosion to remove isolated noise regions with an area smaller than a preset threshold. The spatial range change trajectory of each connected region is recorded in chronological order, and a comprehensive detection report containing the spatial distribution map of the accelerated erosion region and the temporal evolution trend curve is output.

Images