Industrial steel frame point cloud adaptive denoising method and system

Abstract

The application discloses an industrial steel frame point cloud self-adaptive denoising method and system, and belongs to the technical field of three-dimensional point cloud processing. In view of the problems that the industrial steel frame point cloud has bottom weeds, floating noise and non-continuous structure artifacts, the existing technology is not complete in denoising, is easy to damage key structures and depends on manual labeling, the application automatically generates training labels through local geometric attributes, trains an encoder-decoder deep network to output point-level structure confidence, removes various noises through confidence self-adaptive coarse filtering and multi-level fine filtering, guarantees the integrity of components through high-confidence core structure clustering backfilling, and finally outputs pure point cloud through coordinate system alignment and Z-axis layered filtering to purify residual noise. The application does not need manual labeling, is efficient in denoising, has zero damage to core structures, is suitable for complex scenes in industrial fields, and provides high-quality data for steel frame size detection and deformation analysis.

Description

Technical Field

[0001] This invention belongs to the fields of three-dimensional point cloud data processing, computer vision and industrial digital inspection technology, and specifically relates to an adaptive denoising method and system for industrial steel frame point clouds. Background Technology

[0002] With the rapid development and widespread adoption of 3D laser scanning technology, high-precision, high-density point cloud data has become the cornerstone for digital archiving, inspection, and analysis in fields such as architecture, infrastructure, and manufacturing. In the operation, maintenance, inspection, and renovation of industrial steel structure buildings, acquiring 3D point clouds of their surfaces using ground-based or airborne laser scanners is a crucial step for current status mapping, damage identification, model comparison, and load analysis.

[0003] However, in actual industrial field scanning, the raw point cloud data acquired is far from a pure, ideal structural surface. It is generally mixed with a large number of noise points unrelated to the target steel structure, mainly including: Environmental and ground noise: scattered point clouds generated by gravel, tools, and vegetation (weeds) on the scanning site, often clustered at the bottom of the point cloud. Equipment and attachment noise: point clouds generated by scaffolding, temporary supports, and other non-structural objects attached to the steel frame. Instrument and measurement noise: random floating points and outliers generated by the laser scanner itself due to multipath effects, beam divergence, and signal interference. Occlusion and artifacts: due to limited or obstructed scanning angles, point clouds may appear missing or discontinuous in structures that should be continuous (such as the back of a steel beam), forming discontinuous structural segments. These noises and artifacts severely contaminate the point cloud data. If used directly for subsequent processing, it will lead to distorted model reconstruction, increased dimensional measurement errors, and incorrect feature extraction, thus affecting the accuracy of engineering judgments. Therefore, efficient and high-fidelity denoising preprocessing of the raw point cloud is a crucial and highly challenging step in the digital workflow of industrial steel structures.

[0004] Currently, the main technical methods for point cloud denoising include traditional geometric filtering methods and deep learning-based methods, but both have significant limitations when dealing with complex industrial steel frame point clouds.

[0005] Traditional geometric filtering methods, such as statistical outlier removal, radius filtering, and voxelization filtering, typically rely on local spatial statistical characteristics (e.g., distance distribution) of points for global or local thresholding. The main drawback of these methods is their inability to distinguish between structural points and noise points with similar local geometric features. For example, a small bolt head (important structure) and a pebble on the ground (noise) may have similar local density and distribution; traditional methods are prone to mistakenly deleting the former or retaining the latter. Applying the same filtering intensity to different regions of the point cloud and structures of varying importance fails to achieve differentiated processing. Key nodes and the steel frame itself in industrial structures have different tolerances to denoising, requiring different processing strategies. Furthermore, aggressive denoising can easily create unwanted holes or fractures at the edges of real thin-walled members and slender beams, disrupting the structural topological connectivity, which is disastrous for subsequent structural analysis. Traditional geometry-based thresholding methods struggle to effectively separate noise closely integrated with the main structure (e.g., scaffolding nodes attached to steel beams) or discontinuous artifacts with spatial distributions similar to the structure. Deep learning-based methods have brought new approaches to point cloud denoising, enabling more "intelligent" differentiation by learning noise and structural patterns from large amounts of data. However, existing deep point cloud denoising methods face the following challenges in industrial steel frame scenarios: Supervised learning methods require a large number of "noise-clean" point cloud pairs as training data. For specific industrial steel frame scenarios, obtaining accurately paired, sufficient, and high-quality training data is extremely costly, and manually labeling point cloud-level noise is extremely time-consuming and labor-intensive. Furthermore, most general-purpose point cloud denoising networks are designed to pursue overall denoising accuracy without explicitly embedding mechanisms to protect the "continuity of industrial steel frame structures." The network may sacrifice the integrity of key nodes and slender members in exchange for a "cleaner" overall visual effect, which contradicts the requirement of "zero tolerance" for the loss of critical structures in engineering applications.

[0006] In summary, existing technologies have failed to adequately address the challenge of balancing "efficient denoising" and "structural protection" in point cloud denoising of industrial steel frames, as well as the difficulty of model training without manually labeled data. Therefore, there is an urgent need for an innovative method that can automatically learn noise characteristics from limited steel frame data and intelligently, hierarchically, and controllably remove various types of noise while ensuring the geometric and topological integrity of the main steel frame structure with extremely high priority. This invention addresses this technological gap. Summary of the Invention

[0007] How to achieve efficient and accurate noise removal without a large amount of manually labeled data, while absolutely ensuring the geometric integrity and spatial continuity of key load-bearing structures of steel frames (such as main beams and nodes), has long been a core challenge in the field of point cloud denoising for industrial steel frames. The technical problem to be solved by this invention is to provide an adaptive denoising method and system for industrial steel frame point clouds, which addresses the shortcomings of the existing technology. This method effectively filters out noise while maximizing the preservation of the geometric features and spatial continuity of key load-bearing structures such as steel frames, beams, columns, and connection nodes. It provides high-quality and highly complete basic point cloud data for subsequent high-precision measurement of steel frame dimensions, structural deformation analysis, reverse engineering modeling (BIM), and safety assessment. This method is used to solve the technical problem of removing complex mixed noise such as ground scattered points, equipment attachment noise, scanning artifacts, floating points, and structural discontinuities caused by occlusion in point cloud data.

[0008] The present invention adopts the following technical solution: An adaptive denoising method for point clouds of industrial steel frames includes the following steps: S1. Obtain the original point cloud of the industrial steel frame and perform preprocessing and normalization; automatically generate noise / structure binary classification labels based on the local neighborhood geometric attributes of the point cloud to build a training dataset, and train or call the pre-trained point cloud denoising deep neural network model. S2. Input the point cloud of the industrial steel frame to be denoised into the point cloud denoising deep neural network model trained in step S1 and perform forward inference to obtain the confidence that each point belongs to the structural point. S3. Based on the confidence level obtained in step S2, perform adaptive threshold coarse filtering, and remove low-confidence points and retain candidate structure points according to the target point cloud retention ratio set by the user or the preset confidence threshold. S4. Perform multi-level fine filtering on the candidate structure points obtained in step S3, perform density-based clustering analysis on the bottom region points and remove bottom weed noise; perform global density-based clustering analysis on the current candidate structure points and remove discontinuous structure fragments and floating noise clusters; perform local density consistency optimization based on the local neighborhood average distance of the points and remove local sparse outliers. S5. Perform cluster-based structural integrity protection on the point cloud after step S4, filter high confidence points above the structural protection threshold and cluster them, select the top K largest clusters as core structural clusters, and forcibly retain the points in the core structural clusters. S6. Determine the layer boundary in the z-axis direction according to the characteristics of the steel frame structure. Layer the point cloud output in step S5 according to the z-axis coordinate. Perform conditional filtering on each layer of the point cloud to obtain the filtered point cloud of each layer. S7. Combine the filtered point clouds obtained from step S6 to output the denoised industrial steel frame point cloud.

[0009] Preferably, in step S1, the automatic generation of noise / structure binary classification labels specifically involves: A spatial index is constructed on the normalized point cloud, and the K nearest neighbors of each point are searched. The local covariance matrix is ​​calculated by principal component analysis and eigenvalue decomposition is performed to obtain local density features and local linearity features. After normalizing the features, a comprehensive score is calculated. A classification threshold is determined based on the preset noise point ratio. Points with a comprehensive score below the threshold are marked as noise points, and points with a comprehensive score above or equal to the threshold are marked as structure points.

[0010] Preferably, in step S1, the point cloud denoising deep neural network model is an encoder-decoder architecture. The encoder extracts local features through stacked one-dimensional convolution, batch normalization and ReLU activation function, obtains global features through global max pooling and fuses them with local features. The decoder maps the fused features and outputs the structural confidence of each point through convolution and Softmax function.

[0011] Preferably, in step S1, the model training uses a weighted cross-entropy loss function, sets class weights to address the sample imbalance between structure points and noise points, uses the AdamW optimizer and combines gradient clipping to complete the training, and applies Gaussian noise to the sampling point coordinates during the training process for data augmentation.

[0012] Preferably, in step S3, the threshold determination method for the adaptive threshold coarse filtering is as follows: The user can directly specify a confidence threshold between 0 and 1; or the user can set the target point cloud retention ratio, and the system will automatically calculate the dynamic threshold based on the percentile distribution of the confidence level.

[0013] Preferably, in step S4, bottom weed noise is extracted by using a height threshold to extract bottom region points, and then removed by using DBSCAN clustering to remove discrete points and small clusters smaller than the minimum effective cluster size; discontinuous structural fragments and floating noise clusters are removed by statistically analyzing the average size of clusters and removing clusters smaller than the average size by a preset proportion; local sparse outliers are removed by calculating the average distance of K nearest neighbors and removing points larger than a preset percentile distance threshold.

[0014] Preferably, in step S5, the structural protection threshold is higher than the coarse filtering confidence threshold in step S3. After selecting the core structural cluster, the index of the points within the cluster in the original point cloud is obtained, and the corresponding index points are forcibly marked as reserved.

[0015] Preferably, in step S6, the multi-condition hierarchical filtering further includes: Align the point cloud coordinate system with the direction of the main steel frame, then divide the point cloud into multiple layers according to the z-axis coordinate, and filter each layer of point cloud by using corresponding geometric condition combinations.

[0016] Preferably, in step S7, when fusing the point clouds after filtering each layer, the color and intensity information of the original point cloud are assigned to the retained points, and the denoised point cloud data in PCD or PLY format is output.

[0017] Secondly, embodiments of the present invention provide an adaptive denoising system for point clouds of industrial steel frames, comprising: The data module is used to acquire the original point cloud of the industrial steel frame and to preprocess and normalize the original point cloud. The module is used to automatically generate noise / structure binary classification labels based on the local neighborhood geometric attributes of point clouds to build a training dataset, and to train or call a pre-trained point cloud denoising deep neural network model. The inference module is used to perform forward inference on the point cloud denoising deep neural network model obtained by inputting the point cloud of the industrial steel frame to be denoised into the point cloud denoising deep neural network model trained, and to obtain the confidence that each point belongs to the structural point. The coarse filtering module is used to perform adaptive threshold coarse filtering based on the confidence level, and to remove low-confidence points and retain candidate structure points according to the target point cloud retention ratio set by the user or the preset confidence threshold. The refinement filtering module is used to perform multi-level refinement filtering on candidate structure points in sequence, perform density-based clustering analysis on bottom region points and remove bottom weed noise; perform global density-based clustering analysis on current candidate structure points and remove discontinuous structure fragments and floating noise clusters; and perform local density consistency optimization based on the average distance of the local neighborhood of the point and remove local sparse outliers. The structural module is used to filter high-confidence points that are higher than the structural protection threshold and cluster them, select the top K largest clusters as the core structural clusters, and forcibly retain the points in the core structural clusters. The layered filtering module is used to determine the layered boundary in the z-axis direction based on the features of the steel frame structure. It layers the output point cloud according to the z-axis coordinate, performs conditional filtering on each layer of the point cloud, and obtains the filtered point cloud for each layer. The fusion output module is used to fuse the point clouds after filtering at each layer and output the denoised point cloud of the industrial steel frame.

[0018] Thirdly, a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described adaptive denoising method for industrial steel frame point clouds.

[0019] Fourthly, embodiments of the present invention provide a computer-readable storage medium including a computer program, which, when executed by a processor, implements the steps of the above-described adaptive denoising method for industrial steel frame point clouds.

[0020] Fifthly, a chip includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described adaptive denoising method for industrial steel frame point clouds.

[0021] In a sixth aspect, embodiments of the present invention provide an electronic device, including a computer program, which, when executed by the electronic device, implements the steps of the above-described adaptive denoising method for industrial steel frame point clouds.

[0022] Compared with the prior art, the present invention has at least the following beneficial effects: An adaptive denoising method for point clouds of industrial steel frames is proposed, which uses deep learning semantic confidence as the core and combines geometric filtering and structural protection mechanisms to form a closed-loop processing flow. Model training can be completed without manual data annotation, solving the pain point of traditional methods relying on manual parameter tuning. Multi-level filtering removes bottom weeds, floating noise, and sparse outliers, specifically addressing the complex noise of industrial scenarios. A core structure forced preservation mechanism balances denoising intensity and structural integrity, avoiding the accidental deletion of critical components. Multi-condition layered filtering further purifies residual noise, adapting to the linear structural characteristics of steel frames. The entire process is automated, adapting to multi-scenario scanning data in industrial sites, providing high-quality data for subsequent dimensional inspection and deformation analysis, balancing denoising efficiency and structural fidelity, and overcoming the bottlenecks of incomplete denoising and poor structural preservation in traditional methods.

[0023] Furthermore, the steel frame possesses high linearity and high density geometric features. Local geometric attributes are extracted through covariance matrix eigenvalue decomposition, and a weighted comprehensive score is calculated to achieve automatic classification. This completely eliminates manual annotation, significantly reducing the cost of training data construction and solving the problem of time-consuming and labor-intensive annotation in industrial scenarios. Label generation closely matches the structural characteristics of the steel frame, resulting in higher discriminative power and avoiding the poor adaptability issues of general annotation rules. The annotation logic based on local neighborhood geometric attributes adapts to point clouds of steel frames of different specifications and shapes, exhibiting strong generalization. The scoring threshold is adaptively determined according to the noise ratio, eliminating the need for manually preset fixed thresholds and improving annotation accuracy.

[0024] Furthermore, the encoder extracts local geometric features through one-dimensional convolution and captures overall structural features through global max pooling. The fusion of these two methods enhances feature representation capabilities, and the output is a point-level confidence score via Softmax. The network architecture is specifically designed for 3D point clouds, adapting to the disordered and sparse characteristics of point clouds and avoiding the failure of traditional network feature extraction. Local features capture the subtle features of the steel frame components, while global features grasp the overall structural layout. The fusion of these dual features improves the accuracy of noise and structure discrimination. The combination of convolution, batch normalization, and ReLU activation accelerates model convergence and improves inference speed. The point-level confidence score output provides accurate semantic basis for subsequent filtering, replacing the single judgment of traditional geometric thresholds and significantly improving denoising accuracy.

[0025] Furthermore, class weights are set to address the imbalance between structural and noise point samples, gradient clipping ensures training stability, and Gaussian noise enhances model robustness. A weighted loss function addresses the uneven sample distribution problem in automatic labeling, preventing the model from biased towards majority class prediction and improving noise recognition recall. The AdamW optimizer, combined with L2 regularization, prevents overfitting and improves generalization ability. Gradient clipping avoids gradient explosion, ensuring stable convergence during training. Gaussian noise data augmentation simulates real-world scanning noise, improving the model's adaptability to real-world noise and enhancing robustness. The training strategy is fully automated, requiring no manual intervention for tuning.

[0026] Furthermore, based on confidence percentile distribution or user-defined parameters, rapid separation of noise and structure is achieved. Two threshold modes cater to both professional and general users, offering high operational flexibility. Thresholds are automatically calculated based on the retention ratio, eliminating the need for users to possess prior knowledge of noise distribution and lowering the barrier to entry. Confidence thresholds are based on deep learning semantic discrimination, which is more accurate than traditional geometric thresholds, avoiding the accidental deletion of valid structural points. Global coarse filtering quickly removes a large amount of low-confidence noise, reducing the computational load of subsequent fine-tuning filtering and improving overall processing efficiency. Thresholds can be dynamically adjusted to adapt to different denoising intensity requirements.

[0027] Furthermore, relying on DBSCAN density clustering and neighborhood average distance calculation, three types of typical industrial noise are specifically removed. Bottom weed filtering accurately locates ground noise areas, avoiding indiscriminate height clipping that damages the bottom structure; discontinuous structure filtering, based on cluster size discrimination, efficiently removes floating debris and occlusion artifacts without disrupting the continuity of the main structure; local density optimization eliminates sparse outliers, making the point cloud density uniform and improving the accuracy of subsequent modeling; three-level filtering is executed progressively, with each level targeting one type of noise, resulting in more thorough noise removal and no redundant calculations; filtering parameters are adaptively adjusted to adapt to steel frame point clouds of different densities and shapes.

[0028] Furthermore, high-confidence points correspond to the core load-bearing components of the steel frame, with large clusters forming the core structure. Forced backfilling avoids over-filtering. The structural protection threshold is higher than the coarse filtering threshold to prevent critical components from being mistakenly deleted. Core structures are selected based on cluster size, aligning with the structural characteristics of the main beams and nodes of the steel frame as large clusters. A forced retention mechanism for the original index repairs structural errors during the filtering process, ensuring the geometric continuity and topological integrity of the structure. This meets the fidelity requirements of industrial safety inspections for critical components.

[0029] Furthermore, the steel frame has a linear, regular structure, and after coordinate system alignment, layered filtering is more precise, effectively removing residual noise. Aligning the coordinate system with the main body of the steel frame solves the filtering failure problem caused by chaotic coordinate systems during on-site scanning; layered filtering matches the vertical structural features of the steel frame along the Z-axis, with each layer using dedicated filtering conditions for more refined noise removal; layered filtering thoroughly removes residual weeds and useless structural points from previous processes, significantly improving point cloud purity; filtering parameters can be flexibly adjusted to adapt to steel frame structures of different heights and numbers of layers, demonstrating strong generalization; layered processing reduces the computational complexity of single-layer calculations, improves processing speed, and adapts to large-scale point cloud data.

[0030] Furthermore, point cloud attribute information is preserved, and a universal format is adopted for compatibility with subsequent industrial software. Color and intensity attributes are retained to meet the needs of multiple scenarios such as visualization, reverse modeling, and dimensional inspection; the output format is a universal industrial point cloud format, which can be directly imported into BIM and point cloud processing software without format conversion, improving work efficiency; the results of layered filtering are integrated to ensure the integrity and unbrokenness of the point cloud structure; the output data has high purity and structural fidelity; the output process is automated and can generate processing reports.

[0031] It is understood that the beneficial effects of the second to sixth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0032] In summary, the method of this invention integrates the semantic understanding capabilities of deep learning with the accuracy of traditional geometric filtering. Through confidence-guided and structural integrity protection mechanisms, it achieves high-fidelity denoising of point clouds of industrial steel frames. It requires no manual data annotation, can adapt to different noise environments, thoroughly removes noise such as ground weeds and floating points, and absolutely guarantees the geometric continuity of the core load-bearing structure, which is significantly better than traditional single filtering methods.

[0033] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a point cloud image of the original steel frame containing noisy data. Figure 3 A point cloud visualization of the steel frame at two sites containing a small amount of noise; Figure 4 A point cloud visualization of the main body of the steel frame whose coordinate system is not aligned with the main body. Figure 5 A point cloud visualization diagram aligned with the coordinate system of the main steel frame. Figure 6The image shows a visual comparison of the denoising methods of the present invention and the comparative methods from the perspective of a single site. In the image, (a) is the method of the present invention, (b) is statistical filtering, (c) is radius filtering, and (d) is height filtering. Figure 7 The image shows a visual comparison of the denoising methods of the present invention and the comparative methods from two perspectives of the site. In the image, (a) is the method of the present invention, (b) is statistical filtering, (c) is radius filtering, and (d) is height filtering. Figure 8 A schematic diagram of a computer device provided in an embodiment of the present invention; Figure 9 This is a block diagram of a chip provided according to an embodiment of the present invention.

[0035] Among them, 60. Computer equipment; 61. Processor; 62. Memory; 63. Computer program; 600. Electronic device; 610. Processing unit; 620. Storage unit; 6201. Random access memory unit; 6202. Cache memory unit; 6203. Read-only memory unit; 6204. Program / utility; 6205. Program module; 630. Bus; 640. Display unit; 650. Input / output interface; 660. Network adapter; 700. External device. Detailed Implementation

[0036] 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, not all, of the embodiments of the present invention. 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.

[0037] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0038] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0039] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this invention generally indicates that the preceding and following objects have an "or" relationship.

[0040] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0041] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0042] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0043] This invention provides an adaptive denoising method for point clouds of industrial steel frames, which deeply integrates end-to-end deep learning model training with a set of multi-level, configurable post-processing filtering logic to form a complete technical closed loop from "data → model → inference → fine-tuning → protective output". Its core lies in training a structure-aware network in a self-supervised manner, and using its high-dimensional semantic confidence as a guide to drive a series of adaptive filtering operations with adjustable parameters based on traditional geometry and spatial analysis. Finally, it protects the main structure through an intelligent backfilling mechanism. This invention addresses the complex noise interference problems commonly found in industrial site scanning steel frame point clouds, such as bottom ground weeds, floating noise in the air, and discontinuous structures caused by occlusion. Through an end-to-end deep learning model autonomous training, a multi-level adaptive filtering, and a collaborative mechanism of geometric structure integrity protection, it achieves a balance between efficient noise removal and protection of the main steel frame structure. Due to the weed noise and the spatial complexity of the steel frame structure, after the above-mentioned significant noise removal, a multi-condition hierarchical filtering method is introduced based on the characteristics of the steel frame point cloud to remove remaining weeds and special noise. Through the above processing, this invention realizes a complete end-to-end adaptive denoising process from model autonomous training to inference application, combining the semantic understanding capabilities of deep learning with traditional filtering logic based on geometry, space, and density. Experimental results show that the method of the present invention can remove a large amount of noise from the point cloud of the space frame steel structure while maintaining excellent core structure preservation. The output point cloud structure is clear and has good continuity, providing a reliable data foundation for subsequent high-precision applications such as steel frame size detection, deformation analysis, and reverse modeling. The entire method has adjustable parameters and a transparent process, and is suitable for point cloud preprocessing in various complex industrial scenarios.

[0044] Please see Figure 1 The present invention provides an adaptive denoising method for point clouds of industrial steel frames, comprising the following steps: S1, Point Cloud Data Loading To verify the superior performance of the method of this invention on point clouds of complex space frame steel structures, this invention directly uses a laser scanner to scan and obtain a typical point cloud dataset of space frame steel structures. This dataset is representative and includes information such as steel frames and noise points, aiming to make the invention of this paper practical.

[0045] S2, Autonomous Training of End-to-End Deep Learning Denoising Model This step requires no manually labeled point cloud data. Training labels are automatically generated by analyzing the geometric properties of the point cloud itself, and a deep neural network is trained to learn to identify deep features of "structure" and "noise" from the point cloud. The main steps are as follows: S201, Automatic Training Label Generation This step automatically generates labels for the normalized training point cloud by analyzing its geometric properties. Specifically, for each point... Search for it Calculate the local covariance matrix using the nearest neighbor points. :

[0046] in, Let be the centroid of the neighborhood point set. For Perform eigenvalue decomposition to obtain eigenvalues. Then, geometric descriptors are extracted: local linearity is calculated. (Measuring member characteristics) and based on Local density features of nearest neighbor average distance (Measurement of point cloud density):

[0047]

[0048] Next, we will analyze the {L} of the global point cloud. i} and {D i Normalize the results and calculate the overall score:

[0049] Based on a preset noise ratio η, which is a prior hyperparameter rather than a true value that needs to be precisely matched, the scoring threshold is found. .Will Points with certain noise are labeled as noise (0), otherwise they are labeled as structure (1). Thus, we have obtained a training dataset with automatically generated noise / structure labels that are reasonable to a certain extent.

[0050] S202, Network Model Construction and Training This invention constructs a deep network with an encoder-decoder architecture. The encoder consists of four one-dimensional convolutional modules (Conv1d-BN-ReLU) connected in series, with the number of output channels sequentially increasing from [3→32→64→128→256], extracting point features layer by layer. The decoder contains two similar convolutional modules (256→128→64), followed by a convolutional layer with 2 output channels. The network obtains global features through global max pooling and assigns weights accordingly. Fusion with local features:

[0051] In the model training part, this invention creates a SteelDataset. In each iteration, a fixed subset of points and their corresponding labels are randomly sampled from the labeled point cloud dataset, and random Gaussian noise is applied for data augmentation. Supervised training is performed using the automatically generated labels and a weighted cross-entropy loss function. Since "structure points" predominate in the automatic labels, a smaller weight is assigned to the "noise" class (label 0) to encourage the network to avoid easily classifying points as noise. Set a higher weight for the "Structure" class (Label 1). The loss function is:

[0052] Next, train the model for multiple epochs and save the parameters of the model with the best performance.

[0053] S3. Point cloud reasoning and confidence acquisition For any new cloud to be denoising Repeat the above preprocessing and normalization steps to obtain .Will The trained SteelStructureNet is input for forward propagation. The network outputs a structure confidence score for each point in the network. .this It is no longer a score calculated based on simple geometric rules, but a "semantic level" judgment made by the neural network after integrating local and global, multi-level features, which is the core basis for all subsequent processing.

[0054] S4. Confidence-based adaptive threshold coarse filtering This step utilizes semantic confidence for initial, rapid, and global noise removal. The system provides two threshold determination methods for user selection: Method A (Direct Threshold): The user directly specifies the threshold based on experience or required cleanliness. .reserve > The points are used to form a candidate point cloud, Pcandidate. Method B (proportional threshold): The user sets a target overall retention ratio. (If you want to retain 70% of the points, then) =0.7). The system automatically calculates the dynamic threshold.

[0055] reserve > This method is more intuitive, allowing users to indirectly control the noise reduction intensity by adjusting the "compression ratio".

[0056] S5, Multi-level refined post-processing filtering This step performs a three-step progressive processing on the candidate point cloud Pcandidate. Each processing step targets a specific type of noise and can be configured with parameters independently.

[0057] (1) Bottom weed detection: The goal is to remove clusters of weeds deposited on the ground and equipment base. The method is as follows: a) Calculate the height threshold:

[0058] in, It refers to the bottom height ratio.

[0059] b) Extract the bottom point set:

[0060] c) For Perform DBSCAN clustering. Set a small neighborhood radius and a minimum number of points.

[0061] d) Remove all discrete points with a label of -1 from the clustering results, as well as small clusters with fewer than a certain minimum number of points. These are identified as "bottom weeds".

[0062] (2) Discontinuous structure separation: The goal is to remove small fragments suspended in the air that are disconnected from the main structure, as well as artifacts caused by scan occlusion. The method is as follows: a) Perform global DBSCAN clustering on the current point cloud; the parameters can be slightly larger.

[0063] b) Calculate the size of all clusters { }, calculate the average size .

[0064] c) Set a scaling factor Remove all that satisfy the condition. The clusters are identified, along with noise points marked by DBSCAN. This step filters out minor segments based on the assumption that "the main structure must be a large cluster".

[0065] (3) Local density optimization filtering: The goal is to remove those "escaped" points that, although passing the above clustering tests, have abnormally sparse local point distributions, thus making the point cloud density more uniform. The method is as follows: a) For each point p in the current point cloud i Calculate the average distance to its k nearest neighbors (e.g., k=8). .

[0066] b) Analysis { Given the distribution of}, take a higher percentile. As a threshold: .

[0067] c) Remove all that satisfy the condition. These points are too isolated in a local area and are likely noise.

[0068] S6, Intelligent Structural Integrity Protection This is another core innovation of the present invention, which aims to correct the over-filtering that may be caused by step S5. The concept is: "Deep neural networks believe that points with extremely high confidence are highly likely to belong to the real structure; among these high-confidence points, those that form large spatial clusters must be the core skeleton and must be unconditionally protected." Specifically, a high structural protection threshold θ (e.g., 0.7) is set. From the point cloud processed in step S5, all confidence levels are selected. Points greater than θ constitute the high-confidence point set P. high For P high Perform DBSCAN clustering. After clustering, sort the clusters in descending order of size. Then, select the top N largest clusters as the core structure clusters and record the indices of all points in these clusters in the original input point cloud. In the Boolean mask array that ultimately determines whether a point is retained or deleted, force the values ​​of these index positions to be set to "Retain" (True). This is an "OR" operation, meaning that regardless of the reason a point was marked for deletion in steps S4 and S5, it is now "saved".

[0069] S7, Multi-condition Hierarchical Filtering This is another innovation of the present invention, aiming to remove the remaining noise points after filtering in the previous steps. The steps are as follows: First, determine the coordinate system, align the steel frame body with the coordinate system, and save the aligned point cloud data; second, based on the structural characteristics of the steel frame, determine the layer boundary in the Z-axis direction, and divide the point cloud data into multiple layers according to the Z-axis coordinate values; then, apply a specific combination of filtering conditions to each layer of the point cloud for noise reduction.

[0070] S8, Structure Output Each point cloud layer is processed and analyzed independently before multi-layer fusion to obtain a complete denoised point cloud. If the original point cloud contains color or intensity information, it is extracted as well. This data is then recombined to generate a new, clean point cloud file (e.g., PCD, PLY format) and saved. A processing report can also be output, including statistics such as the number of input points, the number of output points, the removal ratio, and the number of points removed in each step.

[0071] In another embodiment of the present invention, an adaptive denoising system for industrial steel frame point clouds is provided. This system can be used to implement the above-mentioned adaptive denoising method for industrial steel frame point clouds. Specifically, the adaptive denoising system for industrial steel frame point clouds includes a data module, a construction module, an inference module, a coarse filtering module, a fine filtering module, a structure module, a hierarchical filtering module, and a fusion output module.

[0072] The data module is used to acquire the original point cloud of the industrial steel frame and to preprocess and normalize the original point cloud. The module is used to automatically generate noise / structure binary classification labels based on the local neighborhood geometric attributes of point clouds to build a training dataset, and to train or call a pre-trained point cloud denoising deep neural network model. The inference module is used to perform forward inference on the point cloud denoising deep neural network model obtained by inputting the point cloud of the industrial steel frame to be denoised into the point cloud denoising deep neural network model trained, and to obtain the confidence that each point belongs to the structural point. The coarse filtering module is used to perform adaptive threshold coarse filtering based on the confidence level, and to remove low-confidence points and retain candidate structure points according to the target point cloud retention ratio set by the user or the preset confidence threshold. The refinement filtering module is used to perform multi-level refinement filtering on candidate structure points in sequence, perform density-based clustering analysis on bottom region points and remove bottom weed noise; perform global density-based clustering analysis on current candidate structure points and remove discontinuous structure fragments and floating noise clusters; and perform local density consistency optimization based on the average distance of the local neighborhood of the point and remove local sparse outliers. The structural module is used to filter high-confidence points that are higher than the structural protection threshold and cluster them, select the top K largest clusters as the core structural clusters, and forcibly retain the points in the core structural clusters. The layered filtering module is used to determine the layered boundary in the z-axis direction based on the features of the steel frame structure. It layers the output point cloud according to the z-axis coordinate, performs conditional filtering on each layer of the point cloud, and obtains the filtered point cloud for each layer. The fusion output module is used to fuse the point clouds after filtering at each layer and output the denoised point cloud of the industrial steel frame.

[0073] This invention provides a terminal device comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, graphics processing units (GPUs), tensor processing units (TPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or function. The processor described in this embodiment can be used in the operation of an adaptive denoising method for industrial steel frame point clouds, including: The process involves acquiring the original point cloud of an industrial steel frame and preprocessing and normalizing it; automatically generating noise / structure binary classification labels based on the local neighborhood geometric attributes of the point cloud to construct a training dataset, and training or calling a pre-trained point cloud denoising deep neural network model; inputting the point cloud of the industrial steel frame to be denoised into the trained point cloud denoising deep neural network model for forward inference to obtain the confidence score of each point belonging to a structure point; performing adaptive threshold coarse filtering based on the obtained confidence scores, removing low-confidence points and retaining candidate structure points according to the user-defined target point cloud retention ratio or a preset confidence threshold; sequentially performing multi-level fine filtering on the obtained candidate structure points, performing density-based clustering analysis on the bottom region points, and removing bottom weed noise; Global density-based clustering analysis is performed on the current candidate structural points to remove discontinuous structural fragments and floating noise clusters; local density consistency optimization is performed based on the average distance of the local neighborhood of each point to remove local sparse outliers; clustering-based structural integrity protection is applied to the processed point cloud, high-confidence points above the structural protection threshold are selected and clustered, and the top K largest clusters are selected as core structural clusters, and points in the core structural clusters are forcibly retained; the z-axis layer boundary is determined according to the characteristics of the steel frame structure, and the output point cloud is layered according to the z-axis coordinate, and conditional filtering is performed on each layer of the point cloud to obtain the filtered point cloud of each layer; the filtered point clouds of each layer are fused together to output the denoised industrial steel frame point cloud.

[0074] Please seeFigure 8 The terminal device is a computer device. In this embodiment, the computer device 60 includes a processor 61, a memory 62, and a computer program 63 stored in the memory 62 and executable on the processor 61. When executed by the processor 61, the computer program 63 implements the adaptive denoising method for industrial steel frame point clouds in this embodiment. To avoid repetition, these details are not elaborated here. Alternatively, when executed by the processor 61, the computer program 63 implements the functions of each model / unit in the adaptive denoising system for industrial steel frame point clouds in this embodiment. To avoid repetition, these details are not elaborated here.

[0075] Computer device 60 can be a desktop computer, laptop, handheld computer, cloud server, or other computing device. Computer device 60 may include, but is not limited to, a processor 61 and a memory 62. Those skilled in the art will understand that... Figure 8 This is merely an example of computer device 60 and does not constitute a limitation on computer device 60. It may include more or fewer components than shown, or combine certain components, or different components. For example, computer device may also include input / output devices, network access devices, buses, etc.

[0076] The processor 61 may be a Central Processing Unit (CPU), or other general-purpose processors, graphics processing units (GPUs), tensor processing units (TPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0077] The memory 62 can be an internal storage unit of the computer device 60, such as a hard disk or memory of the computer device 60. The memory 62 can also be an external storage device of the computer device 60, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on the computer device 60.

[0078] Furthermore, the memory 62 may include both internal storage units of the computer device 60 and external storage devices. The memory 62 is used to store computer programs and other programs and data required by the computer device. The memory 62 can also be used to temporarily store data that has been output or will be output.

[0079] Please see Figure 9 The terminal device is an electronic device 600, which is manifested in the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different platform components (including storage unit 620 and processing unit 610), a display unit 640, etc.

[0080] The storage unit stores program code, which can be executed by the processing unit 610 to perform the steps described in the method section of this specification according to various exemplary embodiments of the present invention. For example, the processing unit 610 can perform actions such as... Figure 1 The steps are shown in the figure.

[0081] Storage unit 620 may include readable media in the form of volatile storage units, such as random access memory (RAM) 6201 and / or cache memory 6202, and may further include read-only memory (ROM) 6203.

[0082] Storage unit 620 may also include a program / utility 6204 having a set (at least one) program module 6205, such program module 6205 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0083] Bus 630 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the multiple bus structures.

[0084] Electronic device 600 can also communicate with one or more external devices 700 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 600, and / or with any device that enables electronic device 600 to communicate with one or more other computing devices (e.g., router, modem). This communication can be performed via input / output interface 650. Furthermore, electronic device 600 can also communicate with one or more networks (e.g., local area network, wide area network, and / or public network, such as the Internet) via network adapter 660. Network adapter 660 can communicate with other modules of electronic device 600 via bus 630. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage platforms.

[0085] This invention also provides a storage medium, specifically a computer-readable storage medium, which is a memory device in a terminal device for storing programs and data. It is understood that the computer-readable storage medium here can include both built-in storage media in the terminal device and extended storage media supported by the terminal device; it can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which can be one or more computer programs (including program code). More specific examples of the computer-readable storage medium include: an electrical connection with one or more wires, a portable disk, a hard disk, random access memory, read-only memory, erasable programmable read-only memory, optical fiber, portable compact disk read-only memory, optical storage device, magnetic storage device, or any suitable combination thereof.

[0086] Computer-readable storage media also include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium can also be any readable medium other than a readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, radio frequency, etc., or any suitable combination thereof.

[0087] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0088] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the adaptive denoising method for industrial steel frame point clouds in the above embodiments; one or more instructions in the computer-readable storage medium are loaded and executed by the processor in the following steps: The process involves acquiring the original point cloud of an industrial steel frame and preprocessing and normalizing it; automatically generating noise / structure binary classification labels based on the local neighborhood geometric attributes of the point cloud to construct a training dataset, and training or calling a pre-trained point cloud denoising deep neural network model; inputting the point cloud of the industrial steel frame to be denoised into the trained point cloud denoising deep neural network model for forward inference to obtain the confidence score of each point belonging to a structure point; performing adaptive threshold coarse filtering based on the obtained confidence scores, removing low-confidence points and retaining candidate structure points according to the user-defined target point cloud retention ratio or a preset confidence threshold; sequentially performing multi-level fine filtering on the obtained candidate structure points, performing density-based clustering analysis on the bottom region points, and removing bottom weed noise; Global density-based clustering analysis is performed on the current candidate structural points to remove discontinuous structural fragments and floating noise clusters; local density consistency optimization is performed based on the average distance of the local neighborhood of each point to remove local sparse outliers; clustering-based structural integrity protection is applied to the processed point cloud, high-confidence points above the structural protection threshold are selected and clustered, and the top K largest clusters are selected as core structural clusters, and points in the core structural clusters are forcibly retained; the z-axis layer boundary is determined according to the characteristics of the steel frame structure, and the output point cloud is layered according to the z-axis coordinate, and conditional filtering is performed on each layer of the point cloud to obtain the filtered point cloud of each layer; the filtered point clouds of each layer are fused together to output the denoised industrial steel frame point cloud.

[0089] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0090] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0091] Taking actual scanned point cloud data of a space frame steel structure as an example, a method based on deep learning feature extraction and geometric structure preservation is adopted. Through multi-scale geometric feature perception and attention mechanisms, deep features of the point cloud are extracted, and the confidence score of each point belonging to the structure is calculated. Then, point cloud filtering is performed using an adaptive threshold to finally obtain the denoised point cloud. The specific implementation steps are as follows: 1. Point cloud data preparation and loading Example 1 uses actual point cloud data of a space frame steel structure. This data was acquired using a terrestrial 3D laser scanner, possessing a realistic engineering background and complex geometric features, and contains a total of 223,905 points. The point cloud data includes complete steel structure components, including typical structures such as steel frames and connection nodes, while also containing a large number of weed points, free noise points, and discontinuities in the steel frame. The data is stored in PCD format and includes three coordinate dimensions: X, Y, and Z.

[0092] 2. Autonomous Training of End-to-End Deep Learning Denoising Model The core of this step is to automatically generate training labels by analyzing the geometric properties of the point cloud itself, without relying on manual annotation, and to train a deep learning model that can distinguish between "structure" and "noise". First, automatic training labels are generated and points are automatically labeled. Then, a convolutional neural network with an encoder-decoder architecture is built, trained on the generated dataset, and the optimal model parameters are saved.

[0093] 3. Point cloud reasoning and confidence acquisition For any new point cloud, the same preprocessing is performed, and the data is input into the trained model for inference. The model outputs a structural confidence score for each point in the point cloud. This score represents the semantic probability that the point belongs to the real steel frame structure after the network integrates multi-level features. It is the core basis for all subsequent processing steps.

[0094] 4. Confidence-based adaptive threshold coarse filtering This step uses confidence scores for an initial, rapid, and global noise filtering. The system provides two threshold determination methods: direct threshold and proportional threshold. Users can set a target overall retention rate R, and the system will retain the points with the highest confidence scores, representing the top R proportions.

[0095] 5. Multi-level refined post-processing filtering The candidate point cloud after coarse filtering undergoes a three-step progressive processing to precisely filter specific noise types. These steps are: bottom weed detection and removal, separation of discontinuous structures and floating noise, and local density consistency optimization filtering.

[0096] 6. Intelligent structural integrity protection This step aims to correct any "collateral damage" or excessive erosion that the aforementioned filtering steps may have caused to critical structures, and is crucial for ensuring structural integrity. Deep neural networks consider points with extremely high confidence to be highly likely to belong to the true structure; among these high-confidence points, those that can form large spatial clusters must be the core framework of the steel frame and must be unconditionally protected.

[0097] 7. Multi-condition hierarchical filtering This step aims to obtain cleaner steel frame data. For point clouds containing minor weed noise and useless structures, further analysis of the steel frame's spatial characteristics is used to develop a multi-condition hierarchical filtering method to achieve more accurate noise removal.

[0098] 8. Results Output Based on the final layered point cloud, merge the layered point clouds to generate a new, clean industrial steel frame point cloud file (such as PCD format).

[0099] Table 1. Comparison of steel frame point cloud results from various methods and the method presented in this paper from the perspective of the site.

[0100] To verify the effectiveness and superiority of the method proposed in this invention, a comparative experiment was conducted on the adaptive denoising method for industrial steel frame point clouds based on deep learning feature extraction and geometric structure preservation, and three classic mainstream denoising methods, under the same experimental environment. The experiment used the same steel structure point cloud dataset, with a total of 223,905 points in the original point cloud. The comparison methods included statistical filtering, radius filtering, and height filtering. We introduced indicators such as removal rate, steel frame linearity, structural continuity, connection point preservation, and steel frame beam integrity to conduct a quantitative comparative analysis between the method of this invention and traditional methods. This invention uses point clouds from two different sites for qualitative analysis, with detailed quantitative comparison results analyzed using point clouds from one site, as shown in Table 1.

[0101] The most important aspect of this invention is that, after the intelligent structural integrity protection filtering method mentioned above removes some free noise points and a large number of weed points, the steel frame point cloud still contains some scattered weed points and useless structural points that cannot be removed (such as...). Figure 3 As shown), both site point clouds contain these unremovable noise points. Therefore, this invention further removes these points through multi-condition filtering. First, for the steel frame point cloud obtained by the previous filtering, we check whether the coordinate system is aligned with the main direction of the steel frame. Through visual comparison and analysis, we find that the coordinate system is not aligned with the main direction of the steel frame. Although the Z-axis direction is correct, the other two axes deviate from the overall direction of the steel frame (e.g., Figure 4 Therefore, we align the main body of the steel frame with the coordinate system by rotating it (e.g., Figure 5 This facilitates the next step of coordinate viewing. After aligning the directions, the steel frame point cloud is projected to obtain its X, Y, and Z axis coordinate ranges. For the steel frame point cloud from the perspective of the site of this invention, the X-axis range is [-10.373932, -5.454151], the Y-axis range is [1.967180, 6.984233], and the Z-axis range is [-1.473341, 3.121777]. For the Z-axis direction, we divide it into three intervals according to the steel frame structure, that is, the steel frame is divided into three layers. For the first layer, filtering conditions are set. For the projection information of the first layer point cloud on the XOY plane, the X-axis interval, Y-axis interval, and Z-axis interval are set for conditional filtering. The same method is applied to the point cloud data of the second and third layers for conditional filtering. The filtered point clouds are merged to obtain the complete point cloud, refer to Figure 6 As can be seen, the weeds on the ground have been largely removed, as have some of the more unusual and useless structures within the steel frame itself. For the point cloud from the second-view perspective of the site, we adopted the same method, referring to… Figure 7 Noise points that are difficult to remove have also been removed; therefore, through multi-condition hierarchical filtering, we can obtain a relatively cleaner point cloud steel frame data.

[0102] Under controlled experimental conditions, all methods achieved point removal. However, due to the large number of weed points in the steel frame point cloud and the limitations of the parameter settings in the comparison methods, it is impossible to uniformly compare the denoising rates with other indicators. For a fair comparison, we set the four indicators of steel frame linearity, structural continuity, connection point preservation, and steel frame beam integrity to be roughly the same, and compared their denoising effects. Qualitative and quantitative analysis was performed on the output denoised point cloud files. Here, this invention only performs quantitative analysis on the point cloud from one perspective of the site. Experimental results show that the method of this invention has certain advantages compared with other methods, as detailed below: (1) Under the premise of maintaining a comparable level of structural characteristics, the method of the present invention achieves significantly higher noise reduction efficiency: Experimental data clearly show (as shown in Table 1) that, under the premise of ensuring that the four core structural characteristic indicators of industrial steel frames (steel frame linearity, structural continuity, connection point preservation, and steel frame beam integrity) are maintained at a comparable level, the method of the present invention demonstrates excellent noise reduction capability: In terms of maintaining structural continuity, the method of this invention scored 0.6694, which is similar to statistical filtering (0.6486), radius filtering (0.6359), and height filtering (0.6544). This means that while removing noise points, this method, like the comparative methods, can effectively protect the overall structural integrity and topological connections of the industrial steel frame, avoiding structural breakage or loss of key parts.

[0103] In terms of maintaining the linearity of the steel frame, the method of this invention scored 0.1955, which is basically equivalent to statistical filtering (0.1898), radius filtering (0.1906), and height filtering (0.1863). This indicates that the method of this invention has achieved a level of excellence comparable to traditional methods in protecting the edge features and geometry of linear components of industrial steel frames.

[0104] In terms of connection point preservation and steel frame beam integrity, the method of this invention achieved scores of 1.0010 and 0.3702, respectively, which are essentially on par with the scores of the comparative methods (connection point preservation: 1.0003-1.0009; steel frame beam integrity: 0.3626-0.3711). This demonstrates that the method of this invention is no less effective than any comparative method in protecting the geometric integrity of critical structural nodes and local components.

[0105] While maintaining the aforementioned structural features at the same level, the method of this invention exhibits a significant advantage in denoising efficiency: Figure 2 The original point cloud data of the steel frame shows that there are a large number of weed points, free noise points, and discontinuities in the steel frame. The method of this invention effectively removes these noise points while maintaining the steel structure. Figure 6 , Figure 7This is a visualization comparison of the steel frame point cloud data from two different sites, after denoising using the method of this invention and the comparative method, respectively. (See Table 1 and...) Figure 6 The method of this invention achieves a point removal rate of 22.1%, meaning that 49,385 noise points were removed from the original 223,905 points. It also effectively removed a large number of weed points, detached noise points, and points indicating discontinuities in the steel frame, resulting in point cloud data usable for subsequent registration processing. In contrast, statistical filtering and radius filtering achieved removal rates of only 1.7% and 2.1%, respectively. While these methods removed detached noise points, their limitations in parameter settings prevented them from effectively removing some continuous points in the steel frame. Furthermore, for weed-noise point clouds... Figure 6 , 7 It can be seen that both methods exhibit very poor denoising performance from the perspectives of the two sites, essentially lacking effective weed removal capabilities. For the one-point cloud at each site, while height filtering achieved a removal rate of 19.6%, essentially eliminating weeds completely, it performed poorly for points with free noise and discontinuities in the steel frame. Figure 6 Compared to the original steel frame point cloud data, the original data did not remove most of the detached noise points and discontinuous points in the steel frame, showing a significant difference from the method of this invention. Furthermore, while removing noise using the method of this invention, a comparison with the original steel frame point cloud reveals that the linearity and structure of the steel frame are preserved remarkably well, achieving the goal of maintaining features while removing noise. These comparative results strongly demonstrate that, under the same level of structural feature protection, the method of this invention can identify and remove more weed noise points, achieving a breakthrough in noise reduction efficiency.

[0106] (2) This invention resolves the technical contradiction between the traditional methods' "weak denoising ability" and "passive and conservative structural protection": Traditional denoising methods face an irreconcilable technical contradiction when processing point clouds of industrial steel frames: to improve denoising intensity, structural features are often sacrificed; to maintain structural integrity, the denoising effect is negligible. The method of this invention successfully resolves this contradiction through intelligent feature learning and discrimination mechanisms: From the perspective of the site, comparing the original point cloud data of the grid structure with the data after denoising by various methods, although statistical filtering and radius filtering maintain structural features to a certain extent, their extremely low removal rate (<2.2%) indicates that these two methods are almost unable to effectively perform the denoising task. In practical engineering applications, this means that the point cloud processed by these two methods still contains a large amount of original noise, which cannot meet the requirements of high-quality subsequent processing. Height filtering adopts a simple and direct "low-point clipping" strategy, which can achieve a removal rate of 19.6%, but its technical principle determines that it has a fundamental defect: it cannot distinguish between "weeds at low points" and "effective structures at low points". Therefore, although it has a relatively high removal rate, it is inferior to the method of this invention in terms of preserving structural features (especially structural continuity and steel frame linearity). This indicates that highly filtered denoising comes at the cost of indiscriminate structural damage.

[0107] The method of this invention learns the geometric features and contextual semantics of industrial steel frames through a deep learning model, enabling it to intelligently distinguish between "noise attached to the structural surface" and "the structure itself." Therefore, while maintaining structural features at a level comparable to or even slightly better than comparative methods, it achieves higher denoising efficiency. This demonstrates that the method of this invention does not sacrifice structure for denoising, nor does it protect the structure through conservative processing, but rather achieves a balance between "precise denoising" and "precise shape preservation" through intelligent discrimination.

[0108] (3) The intelligent noise recognition mechanism ensures adaptability to complex industrial scenarios: The noise in the point cloud of industrial steel frames has diverse and complex characteristics: there are both discrete splash points and patches of attached weeds; it is distributed on the surface of the structure and also exists in the gaps between the structure, such as the original steel frame data in this paper (e.g. Figure 2This invention's method, based on a deep learning-based intelligent recognition mechanism, demonstrates strong adaptability to complex scenarios: It effectively identifies noise closely adhering to the structure. Traditional statistical filtering and radius filtering rely on local geometric statistical features to identify outliers, but for weeds adhering to the steel frame surface, they may form a continuous distribution with structural points in their local neighborhood, making them difficult to identify as noise. This invention's method, through training, can identify outliers in these "pseudo-continuous" distributions, thus achieving effective removal. It is highly effective in proactively protecting key structural features. This invention not only identifies noise but also proactively identifies and protects structural features crucial to engineering safety. Comparison of two-view point clouds with the original steel frame data point cloud shows that the main structure is almost perfectly preserved, and the connection points in the experimental data maintain a near-perfect value (1.0010), demonstrating the method's ability to protect key areas such as bolted connections and node plates. Figure 2 The point cloud after denoising using the method of this invention shows that the steel frame beam structure is well preserved, which is difficult to achieve with traditional methods based on simple geometric rules.

[0109] (4) Comprehensive performance verifies the value of technological innovation: Experimental data on the comprehensive removal rate and the four structural feature preservation indicators show that the method of this invention has a clear advantage in comprehensive performance. For point clouds from the perspective of a single site, the removal rate, which represents denoising capability, significantly outperforms statistical filtering (1.7%) and radius filtering (2.1%), and is also superior to height filtering (19.6%). In terms of the four indicators representing structural preservation capability, the method of this invention is basically on par with the comparative method, and achieves the highest scores in both structural continuity and steel frame linearity, two core indicators. This performance of "significantly stronger denoising capability without compromising structural feature protection" proves that the method of this invention does not improve one dimension by sacrificing another, but rather achieves synergistic optimization of multiple performance dimensions through technological innovation.

[0110] (5) The technological innovation value of this invention lies in the fact that it introduces deep learning intelligent discrimination into the field of industrial point cloud processing, changing the traditional method's reliance on fixed rules and human experience, and realizing a paradigm shift from "rule-based filtering" to "understanding-based discrimination". This shift not only brings performance improvement, but also provides a new technical path for the intelligent and automated development of industrial point cloud processing.

[0111] In summary, the present invention provides an adaptive denoising method and system for point clouds of industrial steel frames. While maintaining the key structural features of industrial steel frames, it achieves significantly higher denoising efficiency, solves the technical contradictions that traditional methods have long faced, and provides a reliable and intelligent solution for high-quality processing of steel structure point clouds. It has important engineering application value and broad market prospects in fields such as industrial inspection, digital construction, and intelligent operation and maintenance.

[0112] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0113] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0114] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0115] In the embodiments provided by this invention, it should be understood that the disclosed devices / terminals and methods can be implemented in other ways. For example, the device / terminal embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0116] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0117] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0118] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random-access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0119] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0120] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0121] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0122] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. An adaptive denoising method for point clouds of industrial steel frames, characterized in that, Includes the following steps: S1. Obtain the original point cloud of the industrial steel frame and perform preprocessing and normalization; automatically generate noise / structure binary classification labels based on the local neighborhood geometric attributes of the point cloud to build a training dataset, and train or call the pre-trained point cloud denoising deep neural network model. S2. Input the point cloud of the industrial steel frame to be denoised into the point cloud denoising deep neural network model trained in step S1 and perform forward inference to obtain the confidence that each point belongs to the structural point. S3. Based on the confidence level obtained in step S2, perform adaptive threshold coarse filtering, and remove low-confidence points and retain candidate structure points according to the target point cloud retention ratio set by the user or the preset confidence threshold. S4. Perform multi-level fine filtering on the candidate structure points obtained in step S3, perform density-based clustering analysis on the bottom region points and remove bottom weed noise; perform global density-based clustering analysis on the current candidate structure points and remove discontinuous structure fragments and floating noise clusters; perform local density consistency optimization based on the local neighborhood average distance of the points and remove local sparse outliers. S5. Perform cluster-based structural integrity protection on the point cloud after step S4, filter high confidence points above the structural protection threshold and cluster them, select the top K largest clusters as core structural clusters, and forcibly retain the points in the core structural clusters. S6. Determine the layer boundary in the z-axis direction according to the characteristics of the steel frame structure. Layer the point cloud output in step S5 according to the z-axis coordinate. Perform conditional filtering on each layer of the point cloud to obtain the filtered point cloud of each layer. S7. Combine the filtered point clouds obtained from step S6 to output the denoised industrial steel frame point cloud.

2. The adaptive denoising method for point clouds of industrial steel frames according to claim 1, characterized in that, In step S1, the automatic generation of noise / structure binary classification labels specifically involves: A spatial index is constructed on the normalized point cloud, and the K nearest neighbors of each point are searched. The local covariance matrix is ​​calculated by principal component analysis and eigenvalue decomposition is performed to obtain local density features and local linearity features. After normalizing the features, a comprehensive score is calculated. A classification threshold is determined based on the preset noise point ratio. Points with a comprehensive score below the threshold are marked as noise points, and points with a comprehensive score above or equal to the threshold are marked as structure points.

3. The adaptive denoising method for point clouds of industrial steel frames according to claim 1, characterized in that, In step S1, the point cloud denoising deep neural network model is an encoder-decoder architecture. The encoder extracts local features through stacked one-dimensional convolution, batch normalization and ReLU activation function, obtains global features through global max pooling and fuses them with local features. The decoder maps the fused features and outputs the structural confidence of each point through convolution and Softmax function.

4. The adaptive denoising method for point clouds of industrial steel frames according to claim 1, characterized in that, In step S1, the model training uses the weighted cross-entropy loss function, sets class weights to address the sample imbalance between structure points and noise points, and completes training using the AdamW optimizer combined with gradient clipping. During training, Gaussian noise is applied to the coordinates of the sampling points for data augmentation.

5. The adaptive denoising method for point clouds of industrial steel frames according to claim 1, characterized in that, In step S3, the threshold determination method for the adaptive threshold coarse filtering is as follows: The user can directly specify a confidence threshold between 0 and 1; or the user can set the target point cloud retention ratio, and the system will automatically calculate the dynamic threshold based on the percentile distribution of the confidence level.

6. The adaptive denoising method for point clouds of industrial steel frames according to claim 1, characterized in that, In step S4, bottom weed noise is extracted by using a height threshold to extract bottom region points, and then removed by using DBSCAN clustering to remove discrete points and small clusters smaller than the minimum effective cluster size; discontinuous structural fragments and floating noise clusters are removed by statistically analyzing the average size of clusters and removing clusters smaller than the average size by a preset proportion; local sparse outliers are removed by calculating the average distance of K nearest neighbors and removing points larger than a preset percentile distance threshold.

7. The adaptive denoising method for point clouds of industrial steel frames according to claim 1, characterized in that, In step S5, the structural protection threshold is higher than the coarse filtering confidence threshold in step S3. After selecting the core structural cluster, the index of the points within the cluster in the original point cloud is obtained, and the corresponding index points are forcibly marked as reserved.

8. The adaptive denoising method for point clouds of industrial steel frames according to claim 1, characterized in that, In step S6, the multi-condition hierarchical filtering further includes: Align the point cloud coordinate system with the direction of the main steel frame, then divide the point cloud into multiple layers according to the z-axis coordinate, and filter each layer of point cloud by using corresponding geometric condition combinations.

9. The adaptive denoising method for point clouds of industrial steel frames according to claim 1, characterized in that, In step S7, when fusing the point clouds after filtering each layer, the color and intensity information of the original point cloud are assigned to the retained points, and the denoised point cloud data in PCD or PLY format is output.

10. An adaptive noise reduction system for point clouds of industrial steel frames, characterized in that, include: The data module is used to acquire the original point cloud of the industrial steel frame and to preprocess and normalize the original point cloud. The module is used to automatically generate noise / structure binary classification labels based on the local neighborhood geometric attributes of point clouds to build a training dataset, and to train or call a pre-trained point cloud denoising deep neural network model. The inference module is used to perform forward inference on the point cloud denoising deep neural network model obtained by inputting the point cloud of the industrial steel frame to be denoised into the point cloud denoising deep neural network model trained, and to obtain the confidence that each point belongs to the structural point. The coarse filtering module is used to perform adaptive threshold coarse filtering based on the confidence level, and to remove low-confidence points and retain candidate structure points according to the target point cloud retention ratio set by the user or the preset confidence threshold. The refinement filtering module is used to perform multi-level refinement filtering on candidate structure points in sequence, perform density-based clustering analysis on bottom region points and remove bottom weed noise; perform global density-based clustering analysis on current candidate structure points and remove discontinuous structure fragments and floating noise clusters; and perform local density consistency optimization based on the average distance of the local neighborhood of the point and remove local sparse outliers. The structural module is used to filter high-confidence points that are higher than the structural protection threshold and cluster them, select the top K largest clusters as the core structural clusters, and forcibly retain the points in the core structural clusters. The layered filtering module is used to determine the layered boundary in the z-axis direction based on the features of the steel frame structure. It layers the output point cloud according to the z-axis coordinate, performs conditional filtering on each layer of the point cloud, and obtains the filtered point cloud for each layer. The fusion output module is used to fuse the point clouds after filtering at each layer and output the denoised point cloud of the industrial steel frame.

Images