A self-adaptive matching reconstruction method and system for a deep-sea semi-buried multi-metal nodule

Abstract

The application discloses a kind of deep sea semi-buried multi-metal nodule self-adaptive matching reconstruction method and system.Method comprising: step 1: construct seabed multi-metal nodule dataset, train segmentation model, and classify and segment semi-buried nodule and full naked nodule;Step 2: extract full naked nodule target unit, generate pseudo semi-buried nodule sample, generate corresponding candidate two-dimensional reconstruction shape;Step 3: calculate quantitative evaluation index, generate comprehensive evaluation result, and determine the only optimal reconstruction strategy label;Step 4: input into high-resolution edge perception convolutional neural network model for classification training;Using the trained model to adaptively determine the optimal reconstruction strategy of actual semi-buried nodule target, generate nodule prediction segmentation map.The application breaks the limitation of traditional single geometric reconstruction method, realizes the high-precision adaptive recovery of nodule two-dimensional morphology, and can increase the effective recognition area of nodule target in single frame image by about 9.91%.

Description

Technical Field

[0001] This invention relates to the fields of in-situ detection of deep-sea mineral resources, computer vision, image processing and deep learning, and specifically to an adaptive matching reconstruction method and system for semi-buried polymetallic nodules in the deep sea. Background Technology

[0002] Deep-sea polymetallic nodules are rich in various key strategic metal elements such as manganese, copper, cobalt, and nickel, and are an important deep-sea solid mineral resource for alleviating the shortage of terrestrial mineral resources in the future. Accurately assessing the abundance, coverage, and grain size distribution of seabed polymetallic nodules plays a crucial foundational role in the delineation of deep-sea mining areas, resource estimation, and the development of commercial mining equipment.

[0003] Currently, in-situ image acquisition using underwater robots (such as AUVs and ROVs) equipped with optical cameras has become the most efficient and intuitive technique for obtaining information on the distribution of large-area seabed nodules. However, in the real and complex deep-sea geological environment, due to the long-term influence of bottom currents and sediment cover, polymetallic nodules on the seabed often exhibit a "semi-buried" state. In this state, optical images can only record the local area of ​​the nodules exposed on the sediment surface. If the area of ​​the nodules is directly extracted using traditional image segmentation techniques or existing deep learning models (such as U-Net), only the shape of the exposed part can be obtained, thus seriously underestimating the true coverage area and resource reserves of the nodules. Therefore, there is an urgent need to develop a nodule reconstruction method that can accurately deduce and restore the complete two-dimensional morphology from the local exposed contours.

[0004] In recent years, shape restoration techniques for occluded targets have made some progress, with common geometric restoration methods including the circumscribed rectangle method, morphological dilation correction method, and equivalent circle diameter fitting method. However, the natural morphology of deep-sea polymetallic nodules is extremely complex, encompassing regular ellipsoids, irregular kidney shapes, intergrowth shapes, and even polygonal shapes with uneven surfaces. Existing single geometric reconstruction strategies generally suffer from the limitation of a "one-size-fits-all" approach: ellipse fitting is effective for regular nodules, but easily causes severe area overflow for nodules with depressions or intergrowths; convex hull algorithms cannot characterize tiny topological grooves; and smooth interpolation methods are difficult to adapt to non-convex shapes. More importantly, current image processing frameworks lack an adaptive mechanism that can automatically match the optimal reconstruction algorithm based on the target's own morphological characteristics. After being buried, nodules of different shapes exhibit significant differences in the edge curvature, skeletal topology, and other features of their exposed parts. How to establish a "site-specific" adaptive strategy to select the most physically accurate shape extrapolation scheme for each partially buried nodule with varying shapes has become a core technical bottleneck restricting the development of high-precision resource estimation of seabed nodules. Summary of the Invention

[0005] In view of this, this invention proposes an adaptive matching reconstruction method and system for deep-sea semi-buried polymetallic nodules. During the nodule shape deduction stage, this invention constructs three parallel shape recovery channels, including maximum ellipse, convex hull smoothing, and adaptive inflection point, effectively overcoming the limitations of traditional single geometric methods and significantly improving the reconstruction accuracy and morphological fit for complex deep-sea nodule morphologies (such as conjoined nodules and concave kidney-shaped nodules). Simultaneously, this invention proposes an optimal strategy intelligent matching mechanism based on a high-resolution edge-aware convolutional neural network, which can effectively avoid the contour deduction defects of different reconstruction algorithms when dealing with unsuitable morphologies, fully integrating and utilizing the advantageous strategies of each algorithm under specific incomplete features.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: An adaptive matching reconstruction method for deep-sea semi-buried polymetallic nodules includes: Step 1: Construct a dataset of polymetallic nodules on the seabed, train an improved U-Net segmentation model, and classify and segment semi-buried nodules and fully exposed nodules; Step 2: Extract fully exposed nodule target units from the segmented and labeled image, generate pseudo-semi-buried nodule samples through random burial operation, and for each pseudo-semi-buried nodule sample, use three geometric reconstruction strategies in parallel: elliptic extremum method, convex hull smoothing method and inflection point contouring method to generate corresponding candidate two-dimensional reconstruction shapes. Step 3: Calculate the quantitative evaluation index between each candidate 2D reconstructed shape and the target ground truth, including area ratio, intersection-union ratio and Chamfer distance. Based on this, generate a comprehensive evaluation result for each candidate reconstructed shape and determine the unique optimal reconstruction strategy label. Step 4: Using pseudo-semi-buried tuberculosis samples as input features and the optimal reconstruction strategy label as a supervision signal, input them into a high-resolution edge-aware convolutional neural network model for classification training; use the trained model to adaptively determine the optimal reconstruction strategy for actual semi-buried tuberculosis targets, and superimpose the reconstruction results onto the original segmentation image to generate a tuberculosis prediction segmentation map.

[0007] Step 1 includes: Step 1.1 Extract keyframes from the seabed polymetallic nodule image sequence, perform size normalization and data augmentation, and construct a training dataset; Step 1.2 Annotate the image to distinguish between fully exposed nodules (UBN) and partially buried nodules (SBN), and generate an annotation file containing location, category, and mask information; Step 1.3 Construct a U-Net segmentation network with a pre-trained ResNet50 backbone network and optimize it using a composite loss function consisting of DiceLoss, Focal Loss and Boundary Loss; Step 1.4 Divide the dataset into a training set and a validation set, and train and evaluate the segmentation model.

[0008] The specific procedures for the three geometric reconstruction strategies in step 2 are as follows: (a) Ellipse extreme value method: Extract the closed contour of the outer perimeter of the target and calculate the minimum convex polygon to obtain the convex hull vertex set; search for the point pair with the farthest distance in the convex hull vertex set as the reference major axis of the fitted ellipse, and calculate the ellipse center, semi-major axis and rotation angle; determine the minor semi-axis based on the maximum projection distance from the foreground pixel point set of the pseudo semi-buried target to the reference major axis, thereby generating the reconstructed ellipse shape; (b) Convex hull smoothing method: Extract the outermost closed contour point set of the target, calculate the smallest convex polygon containing the point set and obtain the vertex sequence; use the vertex sequence as control points, use periodic B-spline curves for smooth fitting, and generate the reconstructed shape through polygon filling algorithm; (c) Inflection point contouring method: The binary image is magnified and the contour point sequence is extracted; the local angle is calculated based on the adaptive step size to identify candidate inflection points, and redundant points are eliminated by non-maximum suppression to obtain the final inflection point set; a closed polygon or line segment mask is generated according to the number of inflection points, and the outer boundary is extracted again after fusing it with the original image. The reconstructed shape is generated by B-spline smoothing fitting and size restoration.

[0009] The calculation method for the quantitative evaluation indicators in step 3 is as follows: Area ratio: The ratio of the pixel area of ​​the candidate reconstruction result to that of the target ground truth; Intersection over Union (IoU): The ratio of the area of ​​intersection to the area of ​​union between the candidate reconstruction results and the target ground truth. Chamfer distance: Calculates the bidirectional average nearest Euclidean distance between the candidate reconstructed contour point set and the target ground truth contour point set.

[0010] The rule for determining the optimal reconstruction strategy label in step 4 is as follows: Construct a comprehensive evaluation value, which is composed of the absolute difference between the area ratio and 1, the difference between the intersection-union ratio and 1, and the sum of the Chamfer distance; the strategy with the evaluation value closest to 0 is marked as the optimal reconstruction strategy; if the evaluation values ​​are equal, a weight coefficient is introduced to perform a weighted summation of the area ratio, intersection-union ratio, and Chamfer distance, and the strategy with the weighted score closest to 0 is selected; if they are still equal, the convex hull smoothing method is selected by default.

[0011] Step 4 further includes: constructing a high-resolution edge-aware convolutional neural network model, which retains the full resolution of the image at the input end; using a pseudo-semi-buried nodule binary image as input and the optimal reconstruction strategy label as a supervision signal, and using the cross-entropy loss function for multi-class training, so that it learns the mapping relationship from edge topological features to the optimal reconstruction strategy.

[0012] A two-dimensional reconstruction system for semi-buried polymetallic nodules based on intelligent selection and employing any one of the methods described above, comprising: The tuberculosis segmentation and extraction module is used to classify and segment tubers using an improved U-Net model; The candidate shape generation module is used to generate pseudo-semi-buried samples and execute the elliptic extremum method, convex hull smoothing method and inflection point contouring method in parallel to generate candidate reconstructed shapes. The multidimensional evaluation label determination module is used to determine the optimal reconstruction strategy label based on area ratio, intersection-union ratio, and Chamfer distance; the intelligent decision-making and reconstruction module is used to adaptively select the optimal reconstruction strategy and generate a predicted segmentation map using a high-resolution edge-aware convolutional neural network model.

[0013] An electronic device includes a processor and a memory, wherein the memory stores a computer program that, when executed, implements any of the methods described above.

[0014] A computer-readable storage medium storing computer instructions that, when executed, implement any of the methods described herein.

[0015] The beneficial technical effects of this invention are as follows: This invention discloses an adaptive matching reconstruction method and system for semi-buried polymetallic nodules in deep sea. Compared with existing technologies, this invention employs a convolutional neural network as the mechanism for selecting reconstruction strategies. The network takes a binary mask image of the semi-buried nodule as input and outputs the probability distributions of three geometric reconstruction strategies: elliptic extremum method, convex hull smoothing method, and inflection point contouring method, automatically selecting the optimal strategy accordingly. Simultaneously, this invention constructs three parallel nodule morphology reconstruction channels to perform shape deduction for semi-buried nodules of different morphologies, overcoming the area estimation bias problem of single reconstruction methods when dealing with complex nodule morphologies such as conjoined nodules and concave kidney-shaped nodules. Furthermore, this invention proposes a pseudo-semi-buried training sample generation mechanism based on random burial, used to train the convolutional neural network and verify the applicability of each reconstruction strategy, thereby improving the ability to infer the contours of occluded targets and enhancing the accuracy of restoring local concavities and edge details of nodules. This invention can be applied to fields such as deep-sea polymetallic nodule resource estimation, seabed mineral abundance assessment, and deep-sea environmental exploration. Attached Figure Description

[0016] Figure 1 This is a flowchart of a two-dimensional reconstruction method for semi-buried polymetallic nodules based on multi-strategy adaptive matching, as described in this invention.

[0017] Figure 2 This is a block diagram of the structural modules of the two-dimensional reconstruction system for polymetallic nodules in this invention.

[0018] Figure 3This is a hardware structure block diagram of the electronic device used to execute the two-dimensional reconstruction algorithm in this invention.

[0019] Figure 4 This is a schematic diagram illustrating the effect of pseudo-semi-buried nodule samples in this invention.

[0020] Figure 5 This is a schematic diagram of the elliptic extremum method in this invention.

[0021] Figure 6 This is a schematic diagram of the inflection point topology method in this invention.

[0022] Figure 7 This is a comparison of the two-dimensional morphological reconstruction effects of the same pseudo-buried nodule target in this invention under three different geometric reconstruction strategies.

[0023] Figure 8 This is a schematic diagram of the high-resolution edge-aware convolutional neural network architecture for optimal policy intelligent classification in this invention.

[0024] Figure 9 This is a schematic diagram showing the comparison before and after intelligent reconstruction of a single frame in a seabed image in this invention. Detailed Implementation

[0025] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] This section uses seabed imagery data from Expedition DY41, initially in MP4 format with a resolution of 1920×1080, as well as seabed imagery data from Expedition DY69, also initially in MP4 format with a resolution of 1920×1080. The overall workflow is as follows: Figure 1 As shown, the system modules are as follows Figure 2 As shown, the electronic device hardware is as follows Figure 3 As shown. The process includes: Step 1, constructing a seabed polymetallic nodule dataset and training an improved U-Net segmentation model to perform high-precision classification and segmentation of two main categories of nodules: semi-buried nodules and fully exposed nodules; Step 2, extracting independent, fully exposed nodule target units G from the segmented and labeled images. j A random burial operation is performed on each target unit to generate a pseudo-semi-buried nodule sample P. j For each sample, three independent geometric reconstruction strategies were employed: the elliptic extremum method, the convex hull smoothing method, and the inflection point topology method. 1j C 2j C 3jIn step 2, candidate 2D reconstructed shapes are generated in parallel. Step 3 involves calculating multiple quantitative evaluation metrics between each candidate reconstructed shape and the ground truth value of the target. These metrics include area ratio, intersection-union ratio, and Chamfer distance. Based on these metrics, a comprehensive evaluation result V is generated for each candidate reconstruction model. ij = [R ij , I ij D ij [, and determine the unique optimal reconstruction strategy label L; Step 4, the pseudo semi-buried nodule P generated in step 2] j As input features, the optimal reconstruction strategy label L determined in step 3 is used as a supervision signal and input into the high-resolution edge-aware convolutional neural network model for classification training. The trained model adaptively determines and selects the optimal reconstruction strategy suitable for the semi-buried nodule target extracted by the current actual U-Net segmentation. The reconstruction result based on the optimal strategy is overlaid on the original segmentation image in an α-channel overlay manner to finally generate the nodule prediction segmentation map.

[0027] Step 1 (Dataset Construction and Classification / Segmentation): Train the improved U-Net segmentation model to perform high-precision classification and segmentation of the two main categories of nodules: semi-buried nodules and fully exposed nodules; Step 1.1: Extract keyframes from the seabed polymetallic nodule image sequence and perform preprocessing to construct a training dataset; the preprocessing includes at least size normalization and data augmentation; Step 1.2: Label the preprocessed image to distinguish between two types of targets: fully exposed nodules and partially buried nodules. The corresponding label names are UBN and SBN, and a label file containing target location, category, and mask information is generated. Step 1.3: Construct a U-Net segmentation network, introduce a pre-trained ResNet50 as the backbone feature extraction network, and optimize the network using a composite loss function L. The composite loss function is composed of Dice Loss, Focal Loss and Boundary Loss, as shown in Equation (1): ; In the formula, L Dice To measure the Dice loss of region overlap, L Focal To mitigate the Focal loss due to class imbalance, L Boundary To improve the boundary loss of contour accuracy, α, β, and γ are the corresponding weighting coefficients, where α = 1, β = 1, and γ = 0.5. Step 1.4: Divide the labeled dataset into a training set and a validation set in a 9:1 ratio. Use the training set to train the improved U-Net network and use the validation set to evaluate the model performance to obtain the trained classification and segmentation model.

[0028] Step 2 (Pseudo-sample generation and parallel reconstruction): Extract independent, fully exposed nodule target units G from the segmented and labeled image. j A random burial operation is performed on each target unit to generate a pseudo-semi-buried nodule sample P. j For each sample, three independent geometric reconstruction strategies were employed: the elliptic extremum method, the convex hull smoothing method, and the inflection point topology method. 1j C 2j C 3j In parallel, candidate two-dimensional reconstructed shapes are generated; Step 2.1: Batch output the segmented and labeled images of polymetallic nodules on the seabed using the ISAT tool; Step 2.2: Perform connected component analysis on the segmented and labeled image to extract the discrete foreground pixel set representing the solid structure, and denote its binary mask as M. org (x, y) ∈ {0, 1}; Step 2.3: For each extracted fully exposed nodule target, generate a random burial mask B(x, y) ∈ {0,1}. Replace the foreground pixels of the selected portion with background pixels through logical operations to generate a pseudo-semi-buried nodule target matrix M. pseudo (x, y), the calculation formula is as follows: ; In the formula M org (x, y) is a binary mask for the original fully exposed nodules.

[0029] Step 2.4: For each of the pseudo-semi-buried nodule targets P generated in Step 2.2 j ,like Figure 4 As shown, the following three 2D morphological reconstruction processes are executed in parallel to generate three alternative candidate reconstruction results: (a) Elliptic extremum method (e.g.) Figure 5 As shown), it includes: extracting the outer closed contour of the target, calculating the minimum convex polygon of the contour point set, and obtaining the convex hull vertex set P. hull The search is performed by traversing the vertex set of the convex hull to find the pair of points (p0, p1) with the maximum mutual Euclidean distance, which greatly improves the search efficiency. This search satisfies the following condition: ; The line segment formed by this point pair is used as the reference major axis of the fitted ellipse. The geometric center C(c) of the ellipse is calculated.x , c y and reference semi-axis a: ; Let the rotation angle of the direction vector of the reference major axis be: ; Extract all foreground pixel sets P of the original pseudo-semi-buried target mask For any point q(x, y) ∈ P in the point set... mask Calculate its projection parameter t on the baseline segment p0p1: ; Introducing an interval constraint operator corrects the projection parameters to t' = max(0, min(1, t)), and calculates the shortest Euclidean distance from point q to the line segment. Define the minor semi-axis b of the ellipse as the maximum projected distance: ; Based on parameter (c) x , c y Generate the analytical equation for the reconstructed ellipse using (a, b, θ) and fill in the blanks to obtain the alternative result C. 1j ; (b) The convex hull smoothing method includes: extracting the outermost closed contour point set of the target using connected component analysis; calculating the smallest convex polygon containing this contour point set using a scanning algorithm, and obtaining the convex hull vertex sequence H = {v1, v2, ..., v m The sequence H is considered as control points, and a periodic B-spline curve is used for smoothing. An adaptive smoothing factor s = mλ is introduced, where m is the number of control points and λ is a preset smoothing coefficient. The equation of the fitted curve is: ; In the formula N i,k (u) is the basis function. A smooth contour point set is generated by uniform sampling in the normalized parameter space, and an alternative result C is generated using a polygon filling algorithm. 2j ; (c) Inflection point contouring method (e.g.) Figure 6 As shown), it includes: firstly, using nearest neighbor interpolation to enlarge the original binary image to a preset factor; extracting the connected component contour point sequence C = {p0, p1, ..., p N-1 To identify significant shape inflection points, an adaptive step size k is constructed based on the number of contour points N: ; For any point p on the contour i Define the forward vector and backward vector Calculate the local included angle θ i : ; The included angle θ i Points with an angle less than 120° are marked as candidate inflection points and arranged in ascending order of included angle; then, a non-maximum suppression operation is performed, and an adaptive suppression radius is set. Redundant inflection points in the neighborhood of the retained point r are removed to obtain the final set of inflection points; When the number of inflection points is greater than or equal to 3, calculate its convex hull and generate a closed polygon mask; when the number of inflection points is 2, generate a line segment mask connecting the two points; perform bitwise or fusion operations on the generated mask and the magnified original image. The maximum outer boundary contour is re-extracted from the fused image. After deduplication and downsampling, periodic B-spline smoothing and resampling are performed. Finally, the smoothed and filled image is restored to its original size through region interpolation and then thresholded and binarized to obtain the candidate result C. 3j .

[0030] Step 3 (Multidimensional Quantitative Assessment) includes: Step 3.1: As Figure 7 As shown, for each candidate reconstruction result C ij and its corresponding objective truth value G j Calculate the pixel area of ​​each pixel and then calculate the area ratio R using the following formula. ij ; ; In the formula, Area(·) represents the total number of pixels in the input region. This ratio is used to evaluate the consistency between the reconstructed result and the target ground truth in terms of region size.

[0031] Step 3.2: Perform pixel-level comparison between the candidate reconstruction results and the target ground truth, calculate their intersection area and union area, and calculate the intersection-union ratio I according to the following formula. ij : ; In the formula, C ij ∩ G j C represents the region of overlap between the reconstruction result and the true value. ij ∪ G j This indicates the joint region between the two. This metric is used to quantify the accuracy of the spatial overlap between the reconstruction result and the target ground truth.

[0032] Step 3.3: Extract candidate reconstructed contours C respectively ij and the true value of the target G j The set of pixel edge points, denoted as P Cij With P Gj For point set P Cij For each point in the set P, calculate its distance to the set of points. Gj The Euclidean distance to the nearest point in the middle is calculated, and the average of all distances is taken to obtain the one-way chamfer distance d. Cij → Gj Similarly, calculate d. Gj → Cij The final bidirectional Chamfer distance D ij The arithmetic mean of the two is calculated using the following formula: ; In the formula, x and y represent the coordinate pixels of the corresponding edge point set, ||·||² represents the L2 norm (i.e., Euclidean distance), and |·| represents the number of elements in the point set. This metric is used to evaluate the similarity between the reconstructed contour and the ground truth contour in terms of boundary shape.

[0033] Step 3.4: Reconstruct the contour C for each candidate. ij The three calculated evaluation indicators—area ratio R ij Intersection and Union Ratio I ij and chamfer distance D ij —Combined into a multidimensional feature vector V ij = [R ij , I ij D ij This serves as the basis for a comprehensive quantitative evaluation of the candidate reconstruction model.

[0034] Step 4 (Intelligent Decision Making and Final Reconstruction) includes: Step 4.1: For each pseudo-semi-buried nodule target generated in Step 2, generate the corresponding optimal reconstruction strategy label L based on the evaluation index results of its three candidate reconstruction contours calculated in Step 3; the specific generation rule is: based on R under ideal conditions ij Approaching 1, I ij Approaching 1, D ij Based on the principle of approaching 0, the basic comprehensive evaluation value E of each strategy is constructed, and the calculation formula is as follows; ; The strategy whose baseline evaluation value E is closest to 0 is labeled as the optimal reconstruction strategy L for that objective. If multiple strategies have the same evaluation value, then w is assigned to each indicator. kThe weights are calculated using formula (7) to obtain the weighted composite score E for each strategy. w ; ; In the formula, w1 is 100%, w2 is 80%, and w3 is 60%. Then, the weighted evaluation value E is reselected. w The strategy closest to 0 is used as the optimal label; If multiple strategies still have the same evaluation value after weighted calculation, the system will default to selecting the convex hull smoothing method (label 1) as the final preferred reconstruction scheme for the objective.

[0035] Step 4.2: For each pseudo-semi-buried nodule target, read its two-dimensional buried binary image generated in step 2.3, and pair it with its corresponding optimal reconstruction strategy label L to jointly construct a dataset for training the strategy selection model.

[0036] Step 4.3: Construct a high-resolution edge-aware convolutional neural network model, the structure of which is as follows: Figure 8 As shown, the model consists of an initial convolutional layer (16 channels), edge-aware block 1 (32 channels), edge-aware block 2 (64 channels), edge-aware block 3 (128 channels), and edge-aware block 4 (256 channels), followed by an adaptive pooling layer and a fully connected layer (256 units), with a final output layer dimension of 3. The input retains the full resolution of the image to extract basic morphology and minute inflection point features. The model is trained using the dataset constructed in step 4.2, with the aforementioned optimal reconstruction strategy label L as the supervision signal. The multi-classification error is calculated using the cross-entropy loss function, and the network parameters are optimized through backpropagation. This allows the model to learn the mapping relationship between the edge topological features of the semi-buried nodules and the optimal reconstruction strategy, resulting in a trained intelligent decision-making classification model.

[0037] Step 4.4: Input the binary image of the semi-buried nodule target extracted by the improved U-Net segmentation in the real scene into the trained intelligent decision classification model. The model forward propagates to predict the probability of each strategy and outputs the one with the highest probability as the optimal two-dimensional reconstruction strategy for the target.

[0038] Step 4.5: The system invokes the optimal reconstruction strategy indicated by the model (one of the three reconstruction methods) to extrapolate the complete shape of the target area of ​​the current semi-buried nodule, and then overlays it with the original segmented image through α-channel blending to finally generate a nodule prediction segmentation map containing the complete restored reconstruction contour. As shown in Table 1, for a single frame image in the seabed image, after intelligent reconstruction using the method described in this invention, the number of pixels in the nodule region increased by 9.91% compared to the original segmentation result. The reconstruction results of multimetallic nodules in the seabed image are as follows. Figure 9As shown, the present invention has a significant effect on restoring the complete two-dimensional morphology of nodules obscured by sediments and completing edge details. The reconstructed outline is natural and smooth, and has a very high degree of consistency with the physical truth value of the nodules in the fully exposed state.

[0039] Table 1. Comparison of multi-metallic nodule pixel calculations in single-frame images before and after reconstruction. Original segmented image 31766 16786 48552 Reconstructed image 31049 22311 53360 9.91%

[0040] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, all of which fall within the protection scope of the present invention. The protection scope of the present invention is defined by the appended claims and any equivalent technical solutions.

Claims

1. An adaptive matching reconstruction method for deep-sea semi-buried polymetallic nodules, characterized in that, include: Step 1: Construct a dataset of polymetallic nodules on the seabed, train an improved U-Net segmentation model, and classify and segment semi-buried nodules and fully exposed nodules. Step 2: Extract fully exposed nodule target units from the segmented and labeled image, generate pseudo-semi-buried nodule samples through random burial operation, and for each pseudo-semi-buried nodule sample, use three geometric reconstruction strategies in parallel: elliptic extremum method, convex hull smoothing method and inflection point contouring method to generate corresponding candidate two-dimensional reconstruction shapes. Step 3: Calculate the quantitative evaluation index between each candidate 2D reconstructed shape and the target ground truth, including area ratio, intersection-union ratio and Chamfer distance. Based on this, generate a comprehensive evaluation result for each candidate reconstructed shape and determine the unique optimal reconstruction strategy label. Step 4: Using pseudo-semi-buried tuberculosis samples as input features and the optimal reconstruction strategy label as a supervision signal, input them into a high-resolution edge-aware convolutional neural network model for classification training; use the trained model to adaptively determine the optimal reconstruction strategy for actual semi-buried tuberculosis targets, and superimpose the reconstruction results onto the original segmentation image to generate a tuberculosis prediction segmentation map.

2. The method according to claim 1, characterized in that, Step 1 includes: Step 1.1 Extract keyframes from the seabed polymetallic nodule image sequence, perform size normalization and data augmentation, and construct a training dataset; Step 1.2 Annotate the image to distinguish between fully exposed nodules (UBN) and partially buried nodules (SBN), and generate an annotation file containing location, category, and mask information; Step 1.3 Construct a U-Net segmentation network with a pre-trained ResNet50 backbone network and optimize it using a composite loss function consisting of Dice Loss, Focal Loss and Boundary Loss; Step 1.4 Divide the dataset into a training set and a validation set, and train and evaluate the segmentation model.

3. The method according to claim 1, characterized in that, The specific procedures for the three geometric reconstruction strategies in step 2 are as follows: (a) Ellipse extreme value method: Extract the closed contour of the target and calculate the minimum convex polygon to obtain the convex hull vertex set; search for the pair of points farthest apart in the convex hull vertex set as the reference major axis of the fitted ellipse, and calculate the ellipse center, semi-major axis and rotation angle; The minor axis is determined based on the maximum projection distance from the foreground pixel set of the pseudo-semi-buried target to the reference major axis, thereby generating the reconstructed elliptical shape. (b) Convex hull smoothing method: Extract the outermost closed contour point set of the target, calculate the smallest convex polygon containing the point set and obtain the vertex sequence; use the vertex sequence as control points, use periodic B-spline curves for smooth fitting, and generate the reconstructed shape through polygon filling algorithm; (c) Inflection point contouring method: Enlarge the binary image and extract the contour point sequence; Local angles are calculated based on adaptive step size to identify candidate inflection points. Redundant points are eliminated through non-maximum suppression to obtain the final inflection point set. Closed polygons or line segment masks are generated according to the number of inflection points. After fusing them with the original image, the outer boundary is extracted again. The reconstructed shape is generated by B-spline smoothing and size restoration.

4. The method according to claim 1, characterized in that, The calculation method for the quantitative evaluation indicators in step 3 is as follows: Area ratio: The ratio of the pixel area of ​​the candidate reconstruction result to that of the target ground truth; Intersection over Union (IoU): The ratio of the area of ​​intersection to the area of ​​union between the candidate reconstruction results and the target ground truth. Chamfer distance: Calculates the bidirectional average nearest Euclidean distance between the candidate reconstructed contour point set and the target ground truth contour point set.

5. The method according to claim 1, characterized in that, The rule for determining the optimal reconstruction strategy label in step 4 is as follows: Construct a comprehensive evaluation value, which is composed of the absolute difference between the area ratio and 1, the difference between the intersection-union ratio and 1, and the sum of the Chamfer distance; the strategy with the evaluation value closest to 0 is marked as the optimal reconstruction strategy; if the evaluation values ​​are equal, a weight coefficient is introduced to perform a weighted summation of the area ratio, intersection-union ratio, and Chamfer distance, and the strategy with the weighted score closest to 0 is selected; if they are still equal, the convex hull smoothing method is selected by default.

6. The method according to claim 1, characterized in that, Step 4 further includes: constructing a high-resolution edge-aware convolutional neural network model, which retains the full resolution of the image at the input end; using a pseudo-semi-buried nodule binary image as input and the optimal reconstruction strategy label as a supervision signal, and using the cross-entropy loss function for multi-class training, so that it learns the mapping relationship from edge topological features to the optimal reconstruction strategy.

7. A two-dimensional reconstruction system for semi-buried polymetallic nodules based on intelligent selection and multi-strategy methods, characterized in that, The method described by any one of claims 1-6 comprises: The tuberculosis segmentation and extraction module is used to classify and segment tubers using an improved U-Net model; The candidate shape generation module is used to generate pseudo-semi-buried samples and execute the elliptic extremum method, convex hull smoothing method and inflection point contouring method in parallel to generate candidate reconstructed shapes. The multidimensional evaluation label determination module is used to determine the optimal reconstruction strategy label based on area ratio, intersection-union ratio, and Chamfer distance; the intelligent decision-making and reconstruction module is used to adaptively select the optimal reconstruction strategy and generate a predicted segmentation map using a high-resolution edge-aware convolutional neural network model.

8. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed, implements the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, It stores computer instructions that, when executed, implement the method described in any one of claims 1-6.

Images