Container ship lashing eye plate identification method

Abstract

The application discloses a container ship binding eye plate recognition method, acquires container ship binding bridge original point cloud data, carries out category labeling on eye plate point cloud and non-eye plate area point cloud of the original point cloud data, constructs labeled data set and divides training set and verification set, constructs PointNet network model architecture, loss function and category imbalance strategy, designs original point cloud data sliding window strategy, adopts four-fold cross-validation method to train and evaluate the PointNet network model, and adopts the trained and evaluated PointNet network model to recognize container ship binding eye plates. The application solves the small binding eye plate feature recognition problem in a large binding scene by using an improved PointNet++ neural network model.

Description

Technical Field

[0001] This invention relates to a method for identifying lashing eye plates, and more particularly to a method for identifying lashing eye plates on container ships, belonging to the field of marine technology. Background Technology

[0002] Accuracy testing and verification of lashing systems is a crucial process in the construction of large container ships, directly impacting the lashing and securing results after the ship is delivered and put into operation. Currently, the shipbuilding industry primarily uses on-ship lashing tests and total station inspections for accuracy testing. On-ship lashing tests utilize container test fixtures, with cranes and workers assisting in the stacking and securing of containers one by one to verify whether the installation accuracy of each lashing eyeplate meets the lashing and securing requirements. Total station measurement replaces on-ship lashing tests by obtaining the center point positions of the lashing eyeplates and container corners and analyzing the lashing allowance; however, its single-point measurement process is cumbersome and susceptible to human error. With the rapid development of 3D scanning technology, 3D scanners are increasingly being applied to industrial accuracy inspection processes. Unlike total stations, which require point-by-point measurements, 3D scanners can quickly establish a station, acquiring millions of point cloud data points per second, containing not only 3D coordinate information but also color attributes, comprehensively reflecting the installation accuracy of the lashing eyeplates.

[0003] In recent years, significant progress has been made in the application research of 3D scanning technology in the field of shipbuilding. From the perspective of application objectives, this research can be divided into three main directions: 1) Ship component identification and accuracy detection: Liao Lei [Research on large-scene point cloud data processing algorithm for simulated test container, Dalian University of Technology, 2023] applied 3D scanning technology to container ship test container operations, extracting point cloud data of cargo hold guide rails and stacking cones to achieve fast and accurate simulated test container operations; Xu Shaoqiu et al. [Identification and reconstruction of key lashing features of container ships based on large-scene 3D point clouds, Shipbuilding Engineering, 2022] achieved lashing eye plate position identification and mixed reality reconstruction through point cloud preprocessing and registration technology; Yu Yangfan et al. [Simulated lashing method for container ships based on point cloud registration, Shipbuilding Engineering, 2023] used point cloud registration to solve the hatch cover corner and lashing eye plate pose to achieve lashing allowance extraction.

[0004] While Xu Shaoqiu and Yu Yangfan et al. achieved the extraction of lashing allowance, their research mainly focused on situations where the lashing bridge arrangement was relatively simple, the number of eye plates was small (no more than 100 per lashing bridge), and the distribution was scattered, with only one eye plate typically distributed at a single lashing position. Furthermore, their research relied on the original design data for point cloud cluster segmentation, resulting in poor robustness when the actual installation position deviated significantly from the design position. This paper focuses on the intelligent recognition problem of lashing eye plates for 16,000 TEU and 24,000 TEU ultra-large container ships. Their lashing fasteners feature a multi-eye plate design, with 2-6 eye plates per position and 280-356 eye plates per single-sided lashing bridge, resulting in a dense distribution and significantly increased segmentation difficulty.

[0005] 2) Hull Section Assembly and Installation: Li Chengkun [Research on Identification Method of Docking Features in 3D Point Cloud of Ship Section Joining Surface, Shanghai Jiao Tong University, 2021] developed point cloud filtering and feature registration algorithms to achieve automatic identification of hull section joining surfaces; Kim et al. [Method for estimating management points for dimensional quality management in shipbuilding and offshore plant construction using terrestrial laser scanning and a computer-aided design model, International Journal of Advanced Engineering Informatics, 2024] proposed a method for estimating measurement points from large-scale point cloud data, and used CAD models and ship section point cloud data for registration and alignment, effectively improving the accuracy of point cloud processing.

[0006] 3) Deep Learning Technology for Point Cloud Processing: Qi et al. [Deep learning on point sets for 3D classification and segmentation, Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2017] and [PointNet++: deep hierarchical feature learning on point sets in a metric space, Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach: Curran Associates Inc., 2017] proposed the PointNet and PointNet++ networks, providing a basic framework for direct processing of point cloud data. Subsequently, VoxelNet [VoxelNet: End-to-End Learning for Point CloudBased 3D Object Detection [C]. 2018 IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2018] and RandLA-Net [RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds [C]. 2020 IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2018] and RandLA-Net [RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds [C]. 2020 IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2018] were proposed. Algorithms such as Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 2020, and Point Transformer [Point Transformer [C], 2021 IEEE / CVF International Conference on Computer Vision (ICCV), Montreal, QC, Canada, 2021] have further advanced the development of point cloud semantic segmentation technology.In the field of shipbuilding, Chen Pan et al. [Research on Flange Laser Scanning Point Cloud Segmentation for Intelligent Installation of Ship Shafting System [J]. China Shipbuilding Research, 2023] and Li Rui et al. [Research on Intelligent Recognition Algorithm for Hull Section Joining Surface Components Based on Improved PointNet++ [J]. China Shipbuilding Research, 2024] respectively applied the improved PointNet model to the recognition of ship components.

[0007] Despite the continuous development of point cloud processing technology, its application research in intelligent recognition of ship lashing beams remains insufficient. In particular, there is a lack of systematic solutions for the intelligent recognition of small features in large lashing bridge scenarios.

[0008] Container ship lashing bridges are characterized by their large size and complex structure. A single-sided lashing bridge scan yields point cloud data on the order of 1×10⁸. The lashing eyelets, as key identification targets, are characterized by their small size, large number, and dispersed distribution. Taking a 24,000 TEU container ship as an example, a single-sided lashing bridge measures 60.95×17.42×1.25m, while a single lashing eyelet measures only 0.16×0.108×0.024m, with a total of 356 eyelets, often representing less than 1% of the total points. The original PointNet++ design, intended for semantic segmentation in indoor scenes, is ill-suited for this "large scene, small feature" identification task. Furthermore, the lashing bridge point cloud lacks effective RGB information, making it difficult to directly apply existing deep learning algorithms. Therefore, a specifically designed and improved point cloud recognition algorithm is needed to achieve end-to-end semantic segmentation of the lashing bridge point cloud data. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a method for recognizing lashing eye plates on container ships, which solves the problem of feature recognition of small lashing eye plates in large lashing scenarios.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for identifying lashing eye plates on container ships includes the following steps: S1. Obtain the original point cloud data of the container ship lashing bridge, label the original point cloud data with eye plate point cloud and non-eye plate region point cloud, construct the labeled dataset and divide it into training set and validation set; S2. Construct an improved PointNet network model architecture; S3. Design an improved loss function and class imbalance strategy for the PointNet network model; S4. Design a sliding window strategy for raw point cloud data, decompose the raw point cloud data into a series of overlapping small blocks, perform inference on each block separately, and then merge the results. S5. The improved PointNet network model is trained and evaluated using the four-fold cross-validation method; S6. An improved PointNet network model, after training and evaluation, is used to identify lashing eyeplates on container ships.

[0011] Further, step S1 specifically involves: using a 3D scanner to perform a real-ship scan of the container ship's lashing bridge to obtain raw point cloud data. Each point in the raw point cloud data contains spatial location coordinates (X, Y, Z) and color attributes (R, G, B). Using CloudCompare software, the eyeboard point cloud in the raw point cloud data is labeled as eyeboard, and the non-eyeboard area point cloud in the raw point cloud data is labeled as background. A labeled dataset is constructed, and the labeled dataset is randomly divided into a training set and a validation set at a ratio of 3:1.

[0012] Further, step S2 specifically includes: 2.1 Improved PointNet network model architecture adopts a symmetrical four-layer design, that is, four feature extraction layers and four feature propagation layers, which are connected sequentially in a one-to-one correspondence. 2.2. Based on the geometric characteristics of the eye ligation plate, the sampling radius distribution was optimized by adopting a layer-by-layer increasing sampling radius strategy (0.05, 0.1, 0.2, 0.4), with sampling of (1024, 256, 64, 16) for each layer, while maintaining a fixed 32 neighboring points; 2.3 The feature dimensions are gradually expanded from 32 dimensions in the shallow layer to 512 dimensions.

[0013] Further, step S3 specifically includes: 3.1 Preprocess the CAD model point cloud and the measured point cloud of the eyeplate; 3.2 The shape similarity between the preprocessed CAD model point cloud and the measured point cloud of the eye plate is evaluated from three dimensions: feature value distribution, spatial extensibility and principal direction consistency. 3.3 Determine the comprehensive loss function by combining shape similarity loss and label smoothing cross-entropy loss.

[0014] Furthermore, the comprehensive loss function is: (1) Among them, L total For the total loss, L CAD For shape similarity loss, L BCE The label smoothing cross-entropy loss is used, where α and β are weighting coefficients; Shape similarity loss L CAD for: (2) Among them, L ratioTo measure the difference in the proportion of characteristic values, reflecting whether the eyeplate shape conforms to the standard proportion; L extent To assess the consistency between predicted eyeplate size and standard eyeplate size in terms of measurement range differences; L dir To measure the alignment, ensure that the predicted eyeplate is aligned with the standard eyeplate. Label smoothing cross-entropy loss L BCE for: (3) Among them, y i For real labels, p i For the predicted probability, ϵ is the smoothing factor, C is the number of categories, and N is the total number of points in the point cloud data.

[0015] Further, step S4 specifically includes: 4.1 Point cloud segmentation: Sampling is performed on the original point cloud data using a sliding window of fixed size, with a fixed number of points collected in each sliding window; 4.2 Overlap Sampling: A fixed step size is set between adjacent sliding windows to form an overlapping area; 4.3 Batch Processing: The segmented point cloud data is fed into the PointNet network model in batches for parallel processing; 4.4 Result Mapping: The prediction results of each sliding window are mapped back to the original point cloud coordinate system through index records; 4.5 Result Fusion: For points in overlapping areas, a sequential update strategy is adopted. The prediction results of the post-processed sliding window cover the prediction results of the previously processed sliding window. At the same time, the prediction results are filtered according to the prediction probability value, and the prediction results with confidence above the threshold are retained.

[0016] Further, step S5 specifically includes: 5.1 Experimental environment configuration; 5.2 Progress Evaluation Method: The formula for calculating overall accuracy (OA) is: (4) Among them, TP i As a real example, FP i As a false positive, FN i For false negatives, TN i For a true negative example, N represents the total number of points in the point cloud data; The formula for calculating the average intersection-union ratio (mIoU) is: (5); 5.3. Four models were trained using the 4-fold cross-validation method, and the best-performing model was analyzed. 5.4 Verify the effectiveness of the comprehensive loss function through a control experiment.

[0017] Compared with the prior art, the present invention has the following advantages and effects: 1. This invention constructs a point cloud dataset of lashing scenes based on point cloud data from actual ship scans using a 3D scanner, and solves the problem of feature recognition of small lashing eye plates in large lashing scenes by using an improved PointNet++ neural network model. This invention enables the direct input of complete single-sided lashing bridges into the model for end-to-end prediction without the need for block processing, proving the feasibility of end-to-end prediction inference of the improved PointNet++ model in large scenes. On the validation set, the mean intersection-union ratio (mIoU) reaches 96.61%, the overall accuracy (OA) reaches 99.15%, and the effective recognition rate of eye plates reaches 94.64% on the new test set, basically realizing the intelligent recognition of container ship lashing eye plates. 2. The improvement strategies proposed in this invention include: a four-layer symmetric network architecture design, sampling radius optimization for the characteristics of the eyepiece, geometric prior supervision guided by CAD model, a category-balanced sampling strategy, and a dedicated learning rate scheduling mechanism. These improvements effectively solve the technical problem of identifying small features in large scenes and provide new ideas for the application of point cloud semantic segmentation in the field of ship accuracy inspection. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a container ship lashing plate according to the present invention.

[0019] Figure 2 This is a schematic diagram of the label categories of the labeled dataset of this invention.

[0020] Figure 3 This is a table of the labeled dataset of this invention.

[0021] Figure 4 This is a schematic diagram of the PointNet network model of the present invention.

[0022] Figure 5 This is a schematic diagram of the CAD model point cloud guided learning mechanism of the present invention.

[0023] Figure 6 This is a table showing the experimental environment configuration for this invention.

[0024] Figure 7 This is the accuracy curve of the present invention.

[0025] Figure 8 This is a graph of the average crossover ratio of the present invention.

[0026] Figure 9 This is a graph of the loss function of the present invention.

[0027] Figure 10 This is a comparison chart of the key indicators of this invention.

[0028] Figure 11 This is a schematic diagram of the prediction results of the present invention on an unlabeled point cloud dataset.

[0029] Figure 12 This is a comparison diagram of the model OA of the present invention.

[0030] Figure 13 This is a comparison chart of the changes in the model loss function of this invention.

[0031] Figure 14 This is a comparison chart of the key indicators of the model of this invention. Detailed Implementation

[0032] To illustrate in detail the technical solutions adopted by the present invention to achieve the intended technical objectives, the technical solutions in 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, not all embodiments. Furthermore, the technical means or technical features in the embodiments of the present invention can be replaced without creative effort. The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0033] like Figure 1 As shown, a method for identifying lashing eye plates on container ships includes the following steps: S1. Obtain the original point cloud data of the container ship lashing bridge, label the original point cloud data with eye plate point cloud and non-eye plate region point cloud, construct the labeled dataset and divide it into training set and validation set.

[0034] Constructing a suitable dataset is a crucial foundation for training network models. The original PointNet++ semantic segmentation model was trained and validated using the Stanford 3D Indoor Scenes dataset (S3DIS). This dataset is specifically designed for semantic segmentation of indoor scenes, covering 6 large indoor regions and 13 semantic categories (such as walls, floors, doors, etc.). The S3DIS dataset is characterized by relatively compact scene sizes, large target object sizes, and relatively balanced category distribution; the point cloud size of a single scene typically does not exceed 9×10⁻⁶. 5 point.

[0035] In contrast, this invention addresses the task of fine segmentation of large-scale lashed bridge scenes, where the size of a single lashed eyeplate to be identified is only about one-thousandth of the scene size, exhibiting the typical characteristics of "large scene, small target, and high quantity." The original point cloud size of a single scene exceeds 3×10⁻⁶. 8 The target feature accounts for less than 4% of the total data. This extreme size contrast and severe class imbalance make traditional datasets unsuitable for the needs of this invention, necessitating the construction of a dedicated dataset.

[0036] like Figure 1 As shown, a Trimble X9 3D scanner was used to scan the lashing bridges of a 16,000 TEU container ship to obtain raw point cloud data. Each point contains spatial coordinates (X, Y, Z) and color attributes (R, G, B). This invention uses CloudCompare software for processing and annotation. CloudCompare software is used to annotate the eyeplate point cloud data as eyeplate 1 and the non-eyeplate area point cloud data as background 0, constructing an annotated dataset, as shown below. Figure 2 As shown. The specific method involves using the CloudCompare_Segment tool to segment the bound eyeplate point cloud, then using the CloudCompare_Scalar fields tool to add SF scalar domain labels, and finally saving the data to obtain single eyeplate point cloud data with label information.

[0037] Unlike traditional data preprocessing, this paper does not perform standard normalization on all data, but only aligns the point cloud data to the same coordinate system. This approach preserves the original geometric features of the point cloud, avoids the shape feature loss and excessive point density that may be caused by normalization, and is beneficial for subsequent feature extraction. Without affecting the recognition performance, CloudCompare software is used to uniformly downsample the point cloud, ultimately obtaining a labeled dataset containing nine bridge faces.

[0038] To reliably evaluate model performance, this paper adopts a four-fold cross-validation method, randomly dividing eight labeled datasets into training and validation sets in a 3:1 ratio, and using one unlabeled dataset as an independent test set to verify the model's generalization ability. Figure 3 The dataset information is detailed, including the location of the tying bridge, the number of point clouds, the number of eyeplate points, the number of background points, the proportion of eyeplates, and their uses for each dataset.

[0039] S2. Construct an improved PointNet network model architecture.

[0040] The original PointNet++ semantic segmentation network employs an encoder-decoder architecture, effectively handling unstructured point cloud data. The network primarily consists of a Set Abstraction (SA) layer and a Feature Propagation (FP) layer. The SA layer reduces the point cloud dimensionality through farthest point sampling and constructs local regions using the ball query method, thus improving PointNet for local feature extraction. The FP layer restores the point cloud density through interpolation and fuses multi-scale features to complete the classification task. However, this architecture struggles to directly handle the small feature recognition task on the LashingNet dataset.

[0041] To address the challenge of semantic segmentation with small features in large-scale scenarios, this invention makes significant improvements to the PointNet++ framework, such as... Figure 4 As shown. Specific improvements include: 2.1 Improved PointNet network model architecture adopts a symmetrical four-layer design, namely, four feature extraction layers and four feature propagation layers. The four feature extraction layers and four feature propagation layers are connected in a one-to-one sequence to form a deeper feature extraction and propagation structure.

[0042] 2.2. Based on the geometric characteristics of the eye plate, the sampling radius distribution was optimized by adopting a layer-by-layer increasing sampling radius strategy (0.05, 0.1, 0.2, 0.4), and sampling was performed at each layer (1024, 256, 64, 16), maintaining a fixed 32 neighboring points, which improved the model's ability to perceive small features.

[0043] 2.3 The feature dimensions have been expanded layer by layer from 32 dimensions at the shallow level to 512 dimensions, which enhances the ability to express complex semantics.

[0044] This architecture is specifically designed for the characteristics of bridge tying scenarios, solving the problem that standard networks cannot effectively handle small targets in large scenes. It is particularly suitable for feature recognition tasks such as tying eye plates, which are numerous, small in size, and scattered in distribution.

[0045] S3. Design an improved loss function and class imbalance strategy for the PointNet network model.

[0046] In the LashingNet dataset, the ratio of background points to eyeplate points is extremely unbalanced. Traditional cross-entropy loss functions often lead to overconfidence in the dominant class (background) and under-recognition of the minority classes (eyeplates). To address this issue, this invention introduces a supervision mechanism based on the geometric prior of the CAD model and a class balancing strategy.

[0047] The specific process is as follows: Figure 5 As shown: 3.1 Preprocess the CAD model point cloud and the measured point cloud of the eye plate.

[0048] 3.2 The shape similarity between the preprocessed CAD model point cloud and the measured point cloud of the eye plate is evaluated from three dimensions: feature value distribution, spatial extensibility and principal direction consistency.

[0049] 3.3 Determine the comprehensive loss function by combining shape similarity loss and label smoothing cross-entropy loss.

[0050] This mechanism is particularly suitable for tying bridges because the eyeplate, as a standardized component, has clear geometric features and can be effectively constrained through a CAD model.

[0051] The overall loss function is: (1) Among them, L total For the total loss, L CAD For shape similarity loss, L BCE The label smoothing cross-entropy loss is used, and α and β are weighting coefficients. In this embodiment, α and β are set to 0.7 and 0.3, respectively.

[0052] Shape similarity loss L CAD for: (2) Among them, L ratio To measure the difference in the proportion of characteristic values, reflecting whether the eyeplate shape conforms to the standard proportion; L extent To assess the consistency between predicted eyeplate size and standard eyeplate size in terms of measurement range differences; L dir To measure the alignment, the predicted eyeplate must be aligned with the standard eyeplate. These three dimensions collectively constrain the network to learn and generate eyeplate features that conform to CAD specifications.

[0053] Label smoothing cross-entropy loss L BCE for: (3) Among them, y i For real labels, p i To predict probabilities, ϵ is the smoothing factor, C is the number of classes, and N is the total number of points in the point cloud data. In this embodiment, the smoothing factor ϵ is 0.1, the number of classes C is 2, and the label smoothing cross-entropy loss introduces a smoothing coefficient of 0.1 to convert the original hard labels (0 and 1) into soft labels (0.05 and 0.95). Label smoothing can effectively prevent the model from becoming overconfident and improve generalization ability in ambiguous areas such as the eye plate boundary.

[0054] To address the class imbalance problem, this invention designs a specialized sampling strategy: First, regions containing eyeplate categories are identified, and point cloud blocks are preferentially selected centered on these regions to ensure sufficient target features in the training samples; second, a 6x oversampling mechanism is introduced to dynamically adjust the ratio of eyeplates to background points in the training samples; finally, a 3D cube space-aware sampling and target region tracking mechanism is employed during the data loading stage to give priority to sampling of key regions. These strategies effectively alleviate the problem of the low proportion of bound eyeplates in the original point cloud.

[0055] In terms of network optimization, this invention employs a learning rate strategy specifically for small target feature recognition. The initial learning rate is set to 0.001, and then gradually reduced to 0.000001 using a cosine annealing scheduling mechanism. This avoids oscillations in the later stages of training while maintaining sensitivity to small features. The optimizer chosen is AdamW with a weight decay of 0.00001, which accelerates convergence while controlling the risk of overfitting. These optimization measures significantly improve the model's ability to recognize key small components such as eye ligatures.

[0056] S4. Design a sliding window strategy for the original point cloud data, decompose the original point cloud data into a series of overlapping small blocks, perform inference on each block separately, and then merge the results.

[0057] In point cloud processing for large-scale lashed bridge scenarios, the data volume of a single point cloud is enormous (typically exceeding 1×10⁻⁶). 8 Directly inputting the complete point cloud into the network for processing (at certain levels) can cause memory overflow. To address this issue, this invention designs a specialized sliding window strategy, enabling efficient inference on large point cloud data.

[0058] The core idea of ​​the sliding window strategy is to decompose a large point cloud scene into a series of overlapping small blocks, perform inference on each block separately, and then merge the results. The specific implementation includes the following steps: 4.1 Point cloud segmentation: using a fixed-size sliding window of 1m 3 Sampling is performed on the original point cloud data, with a fixed number of 15,000 points collected in each sliding window.

[0059] 4.2 Overlapping sampling: A fixed step size of 0.6m is set between adjacent sliding windows to form an overlapping area, ensuring the continuity of point cloud features at the boundary and reducing the boundary effect caused by block division.

[0060] 4.3 Batch Processing: The segmented point cloud data is fed into the improved PointNet network model in 16 batches for parallel processing to improve inference efficiency.

[0061] 4.4 Result Mapping: The prediction results of each sliding window are mapped back to the original point cloud coordinate system through index records to ensure accurate correspondence.

[0062] 4.5 Result Fusion: For points in overlapping areas, a sequential update strategy is adopted. The prediction results of the post-processed sliding window cover the prediction results of the previously processed sliding window. At the same time, the prediction results are filtered according to the prediction probability value, and the prediction results with confidence above the threshold are retained to improve the final classification accuracy.

[0063] This strategy not only solves the memory limitation problem, but also reduces segmentation discontinuity through prediction fusion of overlapping regions, enabling the model to efficiently process complete bridging bridge point cloud data and achieve end-to-end prediction.

[0064] S5. The improved PointNet network model is trained and evaluated using the four-fold cross-validation method.

[0065] 5.1 Experimental environment configuration.

[0066] The experiment used a 4-fold cross-validation method, with each fold training for 70 epochs, a batch size of 16, a point cloud block size of 1.0, and 15,000 sampling points. Figure 6 The experimental environment configuration is detailed, and all experiments are conducted under the same conditions.

[0067] 5.2 Progress Evaluation Method: The overall accuracy (OA) is the percentage of correctly classified points for both the eyeplate and background out of the total number of points. The formula for this is: (4) Among them, TP i For true examples (the number of points correctly predicted as class i), FP i For false positives (the number of points incorrectly predicted as class i), FN i For false negatives (the number of missed predictions for class i), TN i N represents the total number of points in the point cloud data, where N is the number of true negative examples (points correctly predicted as non-class i).

[0068] The mean Intersection over Union (mIoU) is the average of the ratios of the intersection and union of the eyeplate and background prediction results with the ground truth labels. The formula is as follows: (5); The advantage of the mean intersection-union ratio (mIoU) is that it is not affected by the number of categories. Even if the number of eye points is very small, their recognition accuracy is still important to the overall task.

[0069] 5.3. Four models were trained using the 4-fold cross-validation method, and the best-performing model was analyzed.

[0070] Figure 7 and Figure 8 The curves showing the changes in accuracy and mean intersection-over-union (mIoU) with the number of iterations are presented. As can be seen from the training curves, both the overall accuracy (OA) and mIoU exhibit a stable upward trend and eventually converge. The accuracy metric reaches its optimal value of 99.15% in the 56th iteration, and the mIoU metric reaches its highest value of 96.61% in the 67th iteration. Compared to the performance of the original PointNet++ in semantic segmentation in indoor scenes (OA 78.5%, mIoU 47.6%), this model significantly improves the recognition ability of small-feature eyeplates in more challenging large-scale tying scenarios, with an OA improvement of 20.65% and a mIoU improvement of 49.01%, validating the effectiveness of the improved strategy of this invention.

[0071] Figure 9 The loss function curves demonstrate the model's good convergence characteristics and generalization ability. The training and validation losses show similar decreasing trends, eventually stabilizing at around 0.03. The two curves remain close and parallel, indicating that the model generalizes well.

[0072] Figure 10 The radar chart shows that the models perform consistently and in a balanced manner across all metrics, demonstrating the robustness of the algorithm in small feature recognition.

[0073] like Figure 11 As shown, on the PC9 dataset (which was not used for training), 265 out of 280 eyeplates were successfully identified, achieving an effective recognition rate of 94.64%, validating the model's generalization ability and practical application value. This result indicates that the model has significant application potential for intelligent identification of lashed eyeplates in practical engineering, and is expected to provide reliable technical support for shipbuilding accuracy inspection.

[0074] 5.4 Verify the effectiveness of the comprehensive loss function through a control experiment.

[0075] To evaluate the effectiveness of the proposed comprehensive loss function, a control experiment was designed to compare the comprehensive loss function with the standard FocalLoss used in the baseline model. Both sets of experiments used the same network structure, learning rate strategy, and training parameter settings, differing only in the choice of loss function.

[0076] like Figure 12 As shown, the model using the comprehensive loss function exhibits faster convergence and higher stability throughout the training process, achieving a high accuracy level in the early stages of training (the first 5 epochs). In contrast, the model using FocalLoss converges more slowly, and its validation set accuracy fluctuates significantly, especially during epochs 20-35. Ultimately, the validation set accuracy of the model using the comprehensive loss function stabilizes at 99.15%, while the FocalLoss model only reaches 97.61%.

[0077] like Figure 13 The loss function curves show that the training and validation losses of the comprehensive loss function model remain relatively close, indicating that the model has good generalization ability. However, the validation loss of the FocalLoss model fluctuates multiple times during training. Although the final loss value is low, this does not translate into higher accuracy, suggesting that FocalLoss may have overemphasized easily separable samples in this task and failed to learn key features sufficiently.

[0078] like Figure 14As shown, the comparison of the mean Intersection over Union (mIoU) further confirms the advantage of the comprehensive loss function. The mIoU value of the comprehensive loss function model is stable at 96.61%, while that of the FocalLoss model is only 90.25%. This difference is particularly important in the recognition of small targets such as eyeplates, demonstrating the significant improvement effect of the comprehensive loss function proposed in this invention on feature learning of small targets.

[0079] Experimental results show that the comprehensive loss function that integrates the geometric prior of the CAD model can effectively improve the recognition performance of small target features such as eye tethers, verifying the effectiveness of the proposed method.

[0080] S6. An improved PointNet network model, after training and evaluation, is used to identify lashing eyeplates on container ships.

[0081] This invention constructs a point cloud dataset for lashing scenes based on point cloud data from actual ship scans using a 3D scanner. It utilizes an improved PointNet++ neural network model to solve the challenge of recognizing small lashing eyeplate features in large lashing scenarios. This invention enables direct input of complete single-sided lashing bridges into the model for end-to-end prediction without block processing, demonstrating the feasibility of end-to-end prediction and inference using the improved PointNet++ model in large scenes. On the validation set, the mean intersection-over-union (mIoU) reaches 96.61%, and the overall accuracy (OA) reaches 99.15%. On a new test set, the effective eyeplate recognition rate reaches 94.64%, essentially achieving intelligent recognition of container ship lashing eyeplates. The proposed improvement strategies include: a four-layer symmetric network architecture design, sampling radius optimization tailored to the characteristics of lashing eyeplates, CAD model-guided geometric prior supervision, a class-balanced sampling strategy, and a dedicated learning rate scheduling mechanism. These improvements effectively solve the technical challenge of recognizing small features in large scenes, providing a new approach for the application of point cloud semantic segmentation in the field of ship precision inspection.

[0082] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A method for identifying lashing eyeplates on container ships, characterized in that... Includes the following steps: S1. Obtain the original point cloud data of the container ship lashing bridge, label the original point cloud data with eye plate point cloud and non-eye plate region point cloud, construct the labeled dataset and divide it into training set and validation set; S2. Construct an improved PointNet network model architecture; S3. Design an improved loss function and class imbalance strategy for the PointNet network model; Step S3 specifically involves: 3.1 Preprocess the CAD model point cloud and the measured point cloud of the eyeplate; 3.2 The shape similarity between the preprocessed CAD model point cloud and the measured point cloud of the eye plate is evaluated from three dimensions: feature value distribution, spatial extensibility and principal direction consistency. 3.3 Determine the comprehensive loss function by combining shape similarity loss and label smoothing cross-entropy loss; The comprehensive loss function is: （1） Among them, L total For the total loss, L CAD For shape similarity loss, L BCE The label smoothing cross-entropy loss is used, where α and β are weighting coefficients; Shape similarity loss L CAD for: （2） Among them, L ratio To measure the difference in the proportion of characteristic values, reflecting whether the eyeplate shape conforms to the standard proportion; L extent To assess the consistency between predicted eyeplate size and standard eyeplate size in terms of measurement range differences; L dir To measure the alignment, ensure that the predicted eyeplate is aligned with the standard eyeplate. Label smoothing cross-entropy loss L BCE for: （3） Among them, y i For real labels, p i To predict probabilities, C is the smoothing factor, C is the number of categories, and N is the total number of points in the point cloud data; S4. Design a sliding window strategy for raw point cloud data, decompose the raw point cloud data into a series of overlapping small blocks, perform inference on each block separately, and then merge the results. S5. The improved PointNet network model is trained and evaluated using the four-fold cross-validation method; S6. An improved PointNet network model, after training and evaluation, is used to identify lashing eyeplates on container ships.

2. The method for identifying lashing plates on container ships according to claim 1, characterized in that: Step S1 specifically involves: using a 3D scanner to perform a real-ship scan of the container ship's lashing bridge to obtain raw point cloud data. Each point in the raw point cloud data contains spatial location coordinates X, Y, Z and color attributes R, G, B. Using CloudCompare software, the eyeboard point cloud in the raw point cloud data is labeled as eyeboard, and the non-eyeboard area point cloud in the raw point cloud data is labeled as background. A labeled dataset is constructed, and the labeled dataset is randomly divided into a training set and a validation set at a ratio of 3:

1.

3. The method for identifying lashing plates on container ships according to claim 1, characterized in that: Step S2 specifically involves: 2.1 Improved PointNet network model architecture adopts a symmetrical four-layer design, that is, four feature extraction layers and four feature propagation layers, which are connected sequentially in a one-to-one correspondence. 2.

2. Optimize the sampling radius distribution based on the geometric characteristics of the eye-binding plate, adopt a layer-by-layer increasing sampling radius strategy, sample each layer separately, and maintain a fixed 32 neighboring points; 2.3 The feature dimensions are gradually expanded from 32 dimensions in the shallow layer to 512 dimensions.

4. The method for identifying lashing plates on container ships according to claim 1, characterized in that: Step S4 specifically involves: 4.1 Point cloud segmentation: Sampling is performed on the original point cloud data using a sliding window of fixed size, with a fixed number of points collected in each sliding window; 4.2 Overlap Sampling: A fixed step size is set between adjacent sliding windows to form an overlapping area; 4.3 Batch Processing: The segmented point cloud data is fed into the improved PointNet network model in batches for parallel processing; 4.4 Result Mapping: The prediction results of each sliding window are mapped back to the original point cloud coordinate system through index records; 4.5 Result Fusion: For points in overlapping areas, a sequential update strategy is adopted. The prediction results of the post-processed sliding window cover the prediction results of the previously processed sliding window. At the same time, the prediction results are filtered according to the prediction probability value, and the prediction results with confidence above the threshold are retained.

5. The method for identifying lashing plates on container ships according to claim 1, characterized in that: Step S5 specifically involves: 5.1 Experimental environment configuration; 5.2 Progress Evaluation Method: The formula for calculating overall accuracy (OA) is: （4） in, TP i As a real example, FP i As a false positive, FN i For false negatives, TN i For a true negative example, N represents the total number of points in the point cloud data; The formula for calculating the average intersection-union ratio (mIoU) is: （5）； 5.

3. Four models were trained using the 4-fold cross-validation method, and the best-performing model was analyzed. 5.4 Verify the effectiveness of the comprehensive loss function through a control experiment.

Images