A tunnel point cloud small target segmentation processing method, system and platform based on curvature perception adaptive network and a storage medium

Abstract

The application discloses a tunnel point cloud small target segmentation processing method, system and platform based on curvature perception adaptive network and a storage medium, and through introduction of a sparse 3D U-Net backbone network, a curvature perception graph structure, an adaptive differentiable sampling mechanism and a geometry-guided Transformer decoder, an end-to-end point cloud segmentation framework is constructed, small targets such as bolt holes and cracks can be efficiently and accurately identified and segmented in massive, sparse and noisy tunnel point clouds, and the precision, robustness and calculation efficiency of small target segmentation in a complex engineering environment are significantly improved, and reliable technical support is provided for tunnel intelligent inspection.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and deep learning technology, specifically relating to a method, system, platform and storage medium for small target segmentation processing of tunnel point clouds based on curvature-aware adaptive networks. Background Technology

[0002] In the safety monitoring and digital transformation of infrastructure in tunnel engineering, accurate detection of small targets such as bolt holes and cracks on the tunnel lining surface is crucial. Three-dimensional point cloud data generated by lidar scanning has become the preferred technology for this task due to its millimeter-level accuracy and non-contact acquisition advantages. However, tunnel point cloud data is characterized by its large scale, sparse distribution, extreme imbalance between the number of background and target points, and dense noise.

[0003] Existing point cloud segmentation methods face significant challenges: deep networks based on point-by-point processing struggle to capture nonlocal geometric relationships; while models based on dense convolutions incur enormous computational costs and cannot effectively focus on the subtle edge features of small targets in noisy environments. Traditional methods often ignore the inherent geometric structural information (such as curvature) in point cloud data that is crucial for identifying small targets, resulting in insufficient accuracy and blurred boundaries when segmenting minute defects.

[0004] Therefore, in order to address the technical problems of insufficient accuracy and blurred boundaries when segmenting minute defects, there is an urgent need to design and develop a method, system, platform, and storage medium for segmenting small targets in tunnel point clouds based on curvature-aware adaptive networks. Summary of the Invention

[0005] To overcome the shortcomings and difficulties of the existing technology, the purpose of this invention is to provide a method, system, platform and storage medium for small target segmentation of tunnel point clouds based on curvature-aware adaptive networks, so as to take into account both high accuracy and high efficiency and make full use of the geometric characteristics of point clouds to solve the problem of small target segmentation.

[0006] The first objective of this invention is to provide a method for segmenting small targets in tunnel point clouds based on curvature-aware adaptive networks; the second objective of this invention is to provide a system for segmenting small targets in tunnel point clouds based on curvature-aware adaptive networks; the third objective of this invention is to provide a platform for segmenting small targets in tunnel point clouds based on curvature-aware adaptive networks; and the fourth objective of this invention is to provide a computer-readable storage medium.

[0007] The first objective of this invention is achieved as follows: the method comprises:

[0008] First data corresponding to the tunnel is created and acquired, and the first data is processed by voxelization through a sparse 3D U-Net backbone network. At the same time, second data corresponding to the first data is generated. The first data is tunnel point cloud data. The second data includes multi-scale global semantic feature data and three-dimensional coordinate feature data.

[0009] A curvature-aware map corresponding to the second data is constructed, and the second data and the third data generated by calculating the deviation of the neighborhood normal vector are fused and processed; wherein, the third data is curvature feature data;

[0010] The curvature-aware map is processed by a differentiable sampling method using an adaptive curvature query network, and fourth data corresponding to the curvature-aware map is generated; wherein, the fourth data consists of a high-resolution sub-map and a low-resolution sub-map.

[0011] The fourth data is enhanced by graph attention network feature enhancement to generate a fifth data corresponding to the fourth data. Based on the fifth data and the second data, a sixth data corresponding to the tunnel is created and generated through geometric adaptive attention mechanism and curvature mask guidance. The fifth data is enhanced multi-resolution feature data, and the sixth data is segmentation prediction result data.

[0012] Furthermore, the sparse 3D U-Net backbone network is processed using sparse convolution operations and computation is performed only on non-empty voxels;

[0013] The sparse 3D U-Net backbone network has an encoder-decoder architecture with skip connections; wherein the encoder and the decoder are symmetrically arranged, the encoder achieves progressive downsampling through multiple layers of sparse convolution, and the decoder achieves progressive upsampling through transposed convolution.

[0014] Each layer of the decoder is fused with the output features of the corresponding layer of the encoder via skip connections.

[0015] Furthermore, the construction of the curvature-aware map corresponding to the second data, and the fusion processing of the second data and the third data generated by calculating the deviation of the neighborhood normal vector, further includes:

[0016] Curvature features are generated by summing and normalizing the normal vector projection deviations within the k-nearest neighbor range, and are corresponding to each node; wherein, the node is each non-empty voxel output by the sparse 3D U-Net.

[0017] Based on the initial node features that integrate global semantic features and curvature features, a graph topology corresponding to the curvature-aware graph is constructed using the k-nearest neighbor algorithm, wherein the weight values ​​of the edges are determined by the spatial proximity and geometric consistency between nodes.

[0018] Furthermore, the step of combining the adaptive curvature query network with differentiable sampling processing of the curvature-aware map to generate fourth data corresponding to the curvature-aware map further includes:

[0019] By combining a multilayer perceptron with shared weights, a seventh data point corresponding to each node in the curvature-sensing graph is created; wherein, the seventh data point is an importance score.

[0020] The seventh data is converted into a normalized sampling probability distribution using a differentiable method. Based on the sampling probability distribution and combined with hierarchical filtering, a node index set corresponding to the probability nodes is created. The node index set includes a high-resolution node index set and a low-resolution node index set.

[0021] Based on the node index set, fourth data corresponding to the curvature-sensing map is induced to be generated.

[0022] Furthermore, the step of enhancing the fourth data through graph attention network features to generate fifth data corresponding to the fourth data, and based on the fifth data and the second data, creating and generating sixth data corresponding to the tunnel through a geometric adaptive attention mechanism and curvature mask guidance, further includes:

[0023] Eighth data corresponding to either the high-resolution subgraph or the low-resolution subgraph is created and generated, and the eighth data is processed by adaptive weight fusion; wherein, the eighth data includes both high-resolution query features and low-resolution query features;

[0024] Based on the third data, a corresponding geometric perception mask is generated, and the mask is multiplied element-wise with the basic attention weights generated through query-key interaction, thereby realizing geometric prior guidance of attention distribution.

[0025] Furthermore, the step of enhancing the fourth data through graph attention network features to generate fifth data corresponding to the fourth data, and based on the fifth data and the second data, creating and generating sixth data corresponding to the tunnel through a geometric adaptive attention mechanism and curvature mask guidance, further includes:

[0026] The enhanced multi-resolution features and the global features extracted by the sparse 3D U-Net backbone network are input into the geometrically adaptive Transformer decoder; wherein, the geometrically adaptive Transformer decoder adopts a three-layer stacked structure, each layer containing a geometrically adaptive attention module, a feedforward network and a residual connection; the feedforward network adopts a two-layer MLP architecture;

[0027] The graph attention network employs a multi-head graph attention layer, incorporating edge weights inherited from the curvature-aware graph as attention biases during feature propagation, and updating node features through two iterative layers.

[0028] The second objective of this invention is achieved as follows: the system is used to implement the tunnel point cloud small target segmentation processing method based on curvature-aware adaptive network, the system comprising:

[0029] The data creation and processing unit is used to create and acquire first data corresponding to the tunnel, and process the first data through a sparse 3D U-Net backbone network voxelization, while generating second data corresponding to the first data; wherein, the first data is tunnel point cloud data; the second data includes multi-scale global semantic feature data and three-dimensional coordinate feature data;

[0030] A data construction and processing unit is used to construct a curvature-aware map corresponding to the second data, and to fuse and process the second data and the third data generated by calculating the deviation of the neighborhood normal vector; wherein, the third data is curvature feature data;

[0031] The first data processing and generation unit is used to combine the adaptive curvature query network with differentiable sampling to process the curvature-aware map and generate fourth data corresponding to the curvature-aware map; wherein, the fourth data is a high-resolution sub-map and a low-resolution sub-map;

[0032] The second data processing and generation unit is used to enhance the fourth data through graph attention network features, generate fifth data corresponding to the fourth data, and based on the fifth data and the second data, create and generate a sixth data corresponding to the tunnel through a geometric adaptive attention mechanism and curvature mask guidance; wherein, the fifth data is enhanced multi-resolution feature data; and the sixth data is segmentation prediction result data.

[0033] Furthermore, the sparse 3D U-Net backbone network is processed using sparse convolution operations and computation is performed only on non-empty voxels;

[0034] The sparse 3D U-Net backbone network has an encoder-decoder architecture with skip connections; wherein the encoder and the decoder are symmetrically arranged, the encoder achieves progressive downsampling through multiple layers of sparse convolution, and the decoder achieves progressive upsampling through transposed convolution.

[0035] Each layer of the decoder is fused with the output features of the corresponding layer of the encoder via skip connections;

[0036] The data construction and processing unit further includes:

[0037] The first generation module is used to generate curvature features corresponding to each node by summing and normalizing the normal vector projection deviations within the k-nearest neighbor range; wherein, the node is each non-empty voxel output by the sparse 3D U-Net;

[0038] The first processing module is used to construct a graph topology structure corresponding to the curvature-aware graph based on the initial node features that integrate global semantic features and the curvature features, using the k-nearest neighbor algorithm, wherein the weight value of the edge is determined by the spatial proximity and geometric consistency between nodes.

[0039] And / or, the first data processing and generation unit further includes:

[0040] The second generation module is used to combine a multilayer perceptron with shared weights to create seventh data corresponding to each node in the curvature perception map; wherein, the seventh data is an importance score.

[0041] The third generation module is used to convert the seventh data into a normalized sampling probability distribution using a differentiable method, and based on the sampling probability distribution, combined with hierarchical filtering processing, to create a node index set corresponding to the probability nodes; wherein, the node index set includes a high-resolution node index set and a low-resolution node index set.

[0042] The fourth generation module is used to induce the generation of fourth data corresponding to the curvature-aware map based on the node index set;

[0043] And / or, the second data processing generation unit further includes:

[0044] The second processing module is used to create and generate eighth data corresponding to the high-resolution subgraph or the low-resolution subgraph respectively, and to adaptively weight-fuse the eighth data; wherein, the eighth data includes high-resolution query features and low-resolution query features;

[0045] The third processing module is used to generate a corresponding geometric perception mask based on the third data, and to perform element-wise multiplication of the mask with the basic attention weights generated through query-key interaction; thereby realizing geometric prior guidance of attention distribution.

[0046] The fourth processing module is used to input the enhanced multi-resolution features and the global features extracted by the sparse 3D U-Net backbone network into the geometrically adaptive Transformer decoder; wherein, the geometrically adaptive Transformer decoder adopts a three-layer stacked structure, each layer containing a geometrically adaptive attention module, a feedforward network and a residual connection; the feedforward network adopts a two-layer MLP architecture;

[0047] The graph attention network employs a multi-head graph attention layer, incorporating edge weights inherited from the curvature-aware graph as attention biases during feature propagation, and updating node features through two iterative layers.

[0048] The third objective of this invention is achieved as follows: it includes a processor, a memory, and a control program for a tunnel point cloud small target segmentation processing platform based on a curvature-aware adaptive network; wherein the control program is executed on the processor, the control program is stored in the memory, and the control program implements the tunnel point cloud small target segmentation processing method based on a curvature-aware adaptive network.

[0049] The fourth objective of this invention is achieved as follows: the computer-readable storage medium stores a control program for a tunnel point cloud small target segmentation processing platform based on a curvature-aware adaptive network, and the control program for the tunnel point cloud small target segmentation processing platform based on a curvature-aware adaptive network implements the tunnel point cloud small target segmentation processing method based on a curvature-aware adaptive network.

[0050] This invention creates and acquires first data corresponding to a tunnel through a method, processes the first data using a sparse 3D U-Net backbone network for voxelization, and simultaneously generates second data corresponding to the first data. The first data is tunnel point cloud data; the second data includes multi-scale global semantic feature data and three-dimensional coordinate feature data. A curvature-aware map corresponding to the second data is constructed, and the second data and third data generated by calculating the deviation of neighborhood normal vectors are fused together. The third data is curvature feature data. The curvature-aware map is processed by differentiable sampling using an adaptive curvature query network, and a fourth data corresponding to the curvature-aware map is generated. The fourth data consists of a high-resolution sub-map and a low-resolution sub-map. The fourth data is enhanced by a graph attention network to generate a fifth data corresponding to the fourth data. Based on the fifth data and the second data, a sixth data corresponding to the tunnel is created and generated using a geometric adaptive attention mechanism and curvature mask guidance. The fifth data is enhanced multi-resolution feature data; the sixth data is segmentation prediction result data; and the method includes a corresponding system, platform, and storage medium to significantly improve the accuracy and robustness of small target segmentation.

[0051] In other words, this invention constructs an end-to-end point cloud segmentation framework by introducing a sparse 3D U-Net backbone network, a curvature-aware graph structure, an adaptive differentiable sampling mechanism, and a geometry-guided Transformer decoder. This framework can efficiently and accurately identify and segment small targets such as bolt holes and cracks in massive, sparse, and noisy tunnel point clouds, improving the accuracy, robustness, and computational efficiency of small target segmentation in complex engineering environments, and providing reliable technical support for intelligent tunnel inspection. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a schematic diagram of the process steps of a tunnel point cloud small target segmentation method based on a curvature-aware adaptive network according to the present invention.

[0054] Figure 2 This is a schematic diagram of the segmentation process of an embodiment of a tunnel point cloud small target segmentation method based on a curvature-aware adaptive network according to the present invention.

[0055] Figure 3This is a schematic diagram of the decoding process of an embodiment of a method for segmenting small targets in tunnel point clouds based on a curvature-aware adaptive network according to the present invention.

[0056] Figure 4 This is a schematic diagram of the tunnel point cloud small target segmentation processing system architecture based on curvature-aware adaptive network according to the present invention.

[0057] Figure 5 This is a schematic diagram of the tunnel point cloud small target segmentation processing platform architecture based on curvature-aware adaptive network according to the present invention.

[0058] Figure 6 This is a schematic diagram of a computer-readable storage medium architecture in one embodiment of the present invention. Detailed Implementation

[0059] To facilitate a clearer understanding of the objectives, technical solutions, and advantages of this invention, the invention will be further described below in conjunction with the accompanying drawings and specific embodiments. Those skilled in the art can easily understand other advantages and effects of this invention from the content disclosed in this specification.

[0060] This invention can also be implemented or applied through other different specific examples, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the spirit of this invention.

[0061] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0062] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Secondly, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0063] Preferably, the tunnel point cloud small target segmentation processing method based on curvature-aware adaptive network of the present invention is applied in one or more terminals or servers. The terminal is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0064] The terminal can be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal can interact with the customer via a keyboard, mouse, remote control, touchpad, or voice control device.

[0065] This invention provides a method, system, platform, and storage medium for segmenting small targets in tunnel point clouds based on curvature-aware adaptive networks.

[0066] like Figure 1 The diagram shown is a flowchart of a tunnel point cloud small target segmentation method based on a curvature-aware adaptive network provided in an embodiment of the present invention.

[0067] In this embodiment, the tunnel point cloud small target segmentation processing method based on curvature-aware adaptive network can be applied to terminals with display functions or fixed terminals. The terminals are not limited to personal computers, smartphones, tablets, desktop computers or all-in-one computers with cameras, etc.

[0068] The tunnel point cloud small target segmentation method based on curvature-aware adaptive networks can also be applied to a hardware environment consisting of a terminal and a server connected to the terminal via a network. The network includes, but is not limited to, wide area networks (WANs), metropolitan area networks (MANs), or local area networks (LANs). The tunnel point cloud small target segmentation method based on curvature-aware adaptive networks in this embodiment can be executed by the server, by the terminal, or by both the server and the terminal.

[0069] For example, for terminals requiring small target segmentation processing of tunnel point clouds based on curvature-aware adaptive networks, the small target segmentation processing function based on curvature-aware adaptive networks provided by the method of this invention can be directly integrated onto the terminal, or a client for implementing the method of this invention can be installed. Furthermore, the method provided by this invention can also run on servers or other devices in the form of a Software Development Kit (SDK), providing an interface for the small target segmentation processing function of tunnel point clouds based on curvature-aware adaptive networks. Terminals or other devices can then implement the small target segmentation processing function of tunnel point clouds based on curvature-aware adaptive networks through the provided interface. The invention will be further described below with reference to the accompanying drawings.

[0070] like Figures 1-3 As shown, this invention provides a method for segmenting small targets in tunnel point clouds based on curvature-aware adaptive networks. The method includes the following steps:

[0071] S01. Create and acquire first data corresponding to the tunnel, and process the first data by voxelization through a sparse 3D U-Net backbone network, while generating second data corresponding to the first data; wherein, the first data is tunnel point cloud data; the second data includes multi-scale global semantic feature data and three-dimensional coordinate feature data;

[0072] S02. Construct a curvature-aware map corresponding to the second data, and fuse the second data and the third data generated by calculating the deviation of the neighborhood normal vector; wherein, the third data is curvature feature data;

[0073] S03. The curvature-aware map is processed by differentiable sampling using an adaptive curvature query network, and fourth data corresponding to the curvature-aware map is generated; wherein, the fourth data consists of a high-resolution sub-map and a low-resolution sub-map.

[0074] S04. The fourth data is enhanced by graph attention network feature enhancement to generate the fifth data corresponding to the fourth data. Based on the fifth data and the second data, a sixth data corresponding to the tunnel is created and generated through geometric adaptive attention mechanism and curvature mask guidance. The fifth data is the enhanced multi-resolution feature data. The sixth data is the segmentation prediction result data.

[0075] The sparse 3D U-Net backbone network is processed using sparse convolution operations and only performs computations on non-empty voxels.

[0076] The sparse 3D U-Net backbone network has an encoder-decoder architecture with skip connections; wherein the encoder and the decoder are symmetrically arranged, the encoder achieves progressive downsampling through multiple layers of sparse convolution, and the decoder achieves progressive upsampling through transposed convolution.

[0077] Each layer of the decoder is fused with the output features of the corresponding layer of the encoder via skip connections.

[0078] The construction of the curvature-aware map corresponding to the second data, and the fusion processing of the second data and the third data generated by calculating the deviation of the neighborhood normal vector, further includes:

[0079] S021. Curvature features corresponding to each node are generated by summing and normalizing the normal vector projection deviations within the k-nearest neighbor range; wherein, the node is each non-empty voxel output by the sparse 3D U-Net.

[0080] S022. Based on the initial node features that integrate global semantic features and the curvature features, a graph topology corresponding to the curvature-aware graph is constructed using the k-nearest neighbor algorithm, wherein the weight values ​​of the edges are jointly determined by the spatial proximity and geometric consistency between nodes.

[0081] The method of combining the adaptive curvature query network with differentiable sampling to process the curvature-aware map and generate fourth data corresponding to the curvature-aware map further includes:

[0082] S031. Using a multilayer perceptron with shared weights, create and generate seventh data corresponding to each node in the curvature perception map; wherein, the seventh data is an importance score;

[0083] S032. The seventh data is converted into a normalized sampling probability distribution using a differentiable method, and based on the sampling probability distribution, combined with hierarchical filtering processing, a node index set corresponding to the probability nodes is created; wherein, the node index set includes a high-resolution node index set and a low-resolution node index set.

[0084] S033. Based on the node index set, induce the generation of fourth data corresponding to the curvature-sensing map.

[0085] The process of enhancing the fourth data through graph attention network features to generate fifth data corresponding to the fourth data, and based on the fifth data and the second data, creating and generating sixth data corresponding to the tunnel through a geometric adaptive attention mechanism and curvature mask guidance, further includes:

[0086] S041. Create and generate eighth data corresponding to the high-resolution subgraph or the low-resolution subgraph respectively, and perform adaptive weight fusion processing on the eighth data; wherein, the eighth data includes high-resolution query features and low-resolution query features;

[0087] S042. Generate a corresponding geometric perception mask based on the third data, and perform element-wise multiplication of the mask with the basic attention weights generated through query-key interaction, thereby realizing geometric prior guidance of attention distribution.

[0088] The process of enhancing the fourth data through graph attention network features to generate fifth data corresponding to the fourth data, and based on the fifth data and the second data, creating and generating sixth data corresponding to the tunnel through a geometric adaptive attention mechanism and curvature mask guidance, further includes:

[0089] S043. The enhanced multi-resolution features and the global features extracted by the sparse 3D U-Net backbone network are input into the geometrically adaptive Transformer decoder; wherein, the geometrically adaptive Transformer decoder adopts a three-layer stacked structure, each layer containing a geometrically adaptive attention module, a feedforward network and a residual connection; the feedforward network adopts a two-layer MLP architecture;

[0090] The graph attention network employs a multi-head graph attention layer, incorporating edge weights inherited from the curvature-aware graph as attention biases during feature propagation, and updating node features through two iterative layers.

[0091] Specifically, in this embodiment of the invention, a curvature-aware adaptive network is provided, which uses a geometry-guided Transformer for small target segmentation in tunnel point clouds. The methods involved are as follows:

[0092] First, a sparse 3D U-Net backbone network is used to voxelize the tunnel point cloud data. Tunnel point cloud segmentation faces challenges of data sparsity and large scale. Traditional dense convolution can easily lead to computational redundancy. Therefore, this embodiment uses a sparse 3D U-Net as the backbone feature extractor, performing operations only on non-empty voxels to efficiently extract multi-scale semantic features (N is the total number of non-empty voxels, 512 is the channel dimension, providing global semantics) and coordinates. (3D coordinates, unit mm, supports geometry injection). This design connects the original point cloud with downstream modules, first inputting the point cloud... (M is the total number of points, The data is voxelized into sparse tensors (voxel size = 5 mm). The encoder then performs progressive downsampling through 5 layers of sparse convolutions, with bottleneck double residual blocks refining the global representation. The decoder performs symmetrical upsampling and fusion with skip connections to recover details. The sparse convolution core calculates the output features through neighborhood weighted summation.

[0093] (1)

[0094] In the formula, Let i be the output vector of the i-th voxel; It is a set of non-empty neighborhood offsets; The weight matrix is ​​a learnable weight matrix; voxel p k The corresponding neighborhood input; This is the offset vector.

[0095] In other words, output features This represents the output vector of the i-th voxel. It is a set of non-empty neighborhood offsets, and the weight matrix can be learned. For a position-independent kernel at offset k, the neighborhood input... From corresponding voxels The bias b is added channel by channel to achieve efficient local aggregation. The decoder then uses transposed convolution to achieve upsampling, gradually recovering high-resolution details from low-resolution encoded features, and further merging with skip connections in the encoding path to avoid information loss.

[0096] (2)

[0097] In the formula, Let be the decoding channel vector of the j-th voxel; For transpose weights; For the upsampled neighborhood offset set; For source coding features;

[0098] Computation depends on the upsampled neighborhood offset set The upsampled output Let j be the decoding channel vector of the j-th voxel, and let the transpose weights be... From the encoder corresponding layer Transpose to share parameters, source coding features Dimensional Matching Transpose Kernel (Low-Resolution Voxel) The output vector, i.e., the sparse convolution of the encoder. In instances at low-resolution layers, resolution restoration is ensured. The computation process first involves concatenating the upsampled features using channel concatenation (Concat). and coding features Combined into higher-dimensional input:

[0099] (3)

[0100] In the formula, The output features of the decoder's layer l; The features are upsampled from the j-th layer. The skip connection features of the encoder's layer l; This ultimately points to global semantic features; Concat(·,·) is a channel-level concatenation operation; Conv() is a convolution operation; BN() is a batch normalization operation; ReLU() is an activation function.

[0101] Dimensionality-reduced convolutional Conv learns cross-layer interactions (compressing redundancy and enhancing correlation), followed by batch normalization (BN) to standardize the feature distribution for stable training, and finally the activation function ReLU introduces nonlinearity to highlight salient patterns.

[0102] Secondly, an Adaptive Curvature Query Network (ACQ-Net) is used for further processing. Curvature-aware graph construction is a fundamental step in ACQ-Net's graph structure building process, aiming to transform unstructured 3D point clouds into structured graphs with rich geometric relationships. The input to this module includes global semantics. Together with 3D coordinates, the system uses feature fusion and graph structure construction to provide high-quality input for subsequent adaptive sampling and graph attention enhancement. Specifically, for each point p, a comprehensive feature representation is assigned: curvature. (Identify high-curvature edges, such as the bevel of a hole), depth (Dimple value, used for screening pore depth) and density (Identifying occlusions or outliers). These cues are calculated by averaging the curvature of the neighborhood normal vector deviations to ensure edge features are captured. Curvature The calculation uses k-nearest neighbors The projection deviations are summed and normalized to highlight high curvature areas, while adapting to density inhomogeneity (through radius adaptation). The curvature formula is as follows:

[0103] (4)

[0104] In the formula, Let be the curvature of point p; p is the current center point; q is a k-nearest neighbor of point p; Let p be the set of k nearest neighbors of point p; Let be the normal vector of point p; Let q be the relative vector from the nearest point q to the center point p; k is the number of nearest neighbors; and d is the average neighborhood distance.

[0105] The input includes the normal vector. (PCA unit vector) and neighborhood mean distance (Normalized scale, parameter k is fixed to balance efficiency and accuracy); relative vector Measured in mm; output is dimensionless. This geometric information and Fusion to generate initial node features It balances global semantics and local geometry, providing a basic input for the construction of curvature-aware maps.

[0106] Next is graph topology construction, based on the initial node features. One of the core innovations of ACQ-Net is the construction of curvature-aware maps. This transforms unstructured point clouds into structured graphs to capture nonlocal geometric dependencies, thereby bridging the global semantic features of U-Net with subsequent graph attention network propagation. The node set V encompasses N voxels, and the edge set... A sparse adjacency matrix is ​​generated using KNN (based on Euclidean distance). The edge weights are designed to consider both spatial proximity and geometric consistency. Integrating multimodal geometric similarity, prioritizing the strengthening of connections in high curvature clusters (through...) (Judgment). High A value close to 1 emphasizes the connectivity of a reliable subgraph, while a low value close to 0 suppresses false points, compensating for the neglect of geometric dependencies in traditional sampling. The edge weight formula is as follows:

[0107] (5)

[0108] In the formula, Let be the final edge weight between point p and point q; The standard deviation of the distance from the Gaussian kernel; Curvature weight; , Let be the curvature eigenvalues ​​of points p and q.

[0109] Input node coordinates (mm unit), curvature Distance scale curvature weight Output Generate a highly sparse adjacency matrix. (Only KNN edge pairs are non-zero), where Measure connection strength, implicitly cluster edge subgraphs with high values, and filter noise. Curvature-aware map. The overall output includes initial node features And the edge adjacency structure (A and metadata), this complete graph directly drives downstream modules.

[0110] Through the above construction process, curvature-sensing map This results in a complete structured representation. This graph structure not only preserves the multimodal feature information of the nodes but also encodes rich geometric relationships through a carefully designed edge weight matrix. This representation provides an ideal data foundation for subsequent processing: the adaptive sampling module can filter nodes based on the importance of their features, while the GAT enhancement module can achieve efficient feature propagation using an edge weight-guided attention mechanism.

[0111] In the adaptive sampling process, the adaptive sampling module is one of the core innovations of this work. Its initial design purpose was to solve the problem of how to replace traditional manual heuristic sampling (such as farthest point sampling) or fixed threshold filtering methods in a differentiable, task-driven manner in large and noisy point cloud scenarios. This module no longer treats sampling as an independent, rigid preprocessing step, but constructs it as a learnable, probabilistic soft selection process. It can dynamically evaluate the importance of each node as a potential target anchor point based on global semantic cues and local geometric characteristics, thereby ensuring that subsequent processing can focus on the most information-rich regions.

[0112] The input to this module is a curvature-sensing map. The initial feature set of each node is determined. In the first stage, namely the sampling probability calculation, the saliency of each node is evaluated through a parameterized function. Specifically, a two-layer multilayer perceptron with shared weights is applied to the initial features of each node. :

[0113] (6)

[0114] In the formula, The importance score for point p; Let be the initial node features of point p; MLP() is a two-layer fully connected network with shared weights; and These are the weights and biases of the two MLP layers, respectively, and σ represents the nonlinear activation function.

[0115] This MLP, as a general function approximator, is optimized to learn a mapping relationship that incorporates U-Net semantics. With geometric properties Mixed characteristics Mapped to an importance score of a scalar .Fraction A higher value indicates that the node is more likely to be located on the edge of a small target of interest or on a critical structure. The score vector for the entire graph is denoted as . .

[0116] In the second stage, differential sampling, M anchor point indices I are selected from N nodes based on the fraction vector s. The Gumbel-Softmax technique is introduced to soften this sampling process. First, the original fraction s is processed through a temperature coefficient... The Softmax function is used to transform a normalized probability distribution. :

[0117] (7)

[0118] In the formula, y i s is the normalized sampling probability of point i; i τ is the original importance score of the i-th point predicted by the previous MLP; τ is the temperature parameter; N is the total number of points in the point cloud.

[0119] The temperature parameter τ (τ > 0) plays a crucial role in this sampling mechanism. As the value of τ approaches zero from a value greater than zero, the output probability distribution y will infinitely approach a vector where only a single position is 1 and all other positions are 0. This characteristic transforms the sampling process from random selection to a deterministic selection mechanism, directly selecting the top M nodes with the highest overall scores. This deterministic mode is particularly advantageous in applications with low noise levels because it enables precise anchor point location. Ultimately, the node with the highest probability is selected from the distribution y. and Each node index constitutes a high-resolution index set. and low-resolution index set The output of this module is not new features, but rather these two index sets. These are used to extract features from the original graph. Two refined subgraphs are induced from the middle. and This induction process preserves the original node features. The sampled subgraph is connected to all edges existing within the index set, along with their weights, ensuring that the sampled subgraph fully inherits the geometric dependencies of the original graph. (High-resolution subgraph) Typically, it contains more nodes, focusing on millimeter-level local details, such as the edge of a bolt hole; while low-resolution subgraphs... With fewer nodes, it aims to cover a broader tunnel background and capture macroscopic curvature distribution patterns. This hierarchical sampling strategy provides multi-scale contextual information for subsequent processing.

[0120] For the GAT (Graph Attention Network) enhancement module, which follows the adaptive sampling module, its core task is to enhance the anchor subgraph provided by adaptive sampling, which has already undergone preliminary screening and purification. Above, it performs efficient and robust graph structure information aggregation. This module iteratively refines node features through stacked multi-head graph attention layers, so that it not only includes the characteristics of the node itself, but also dynamically and weightedly fuses the contextual information of its neighborhood, thereby significantly improving the ability to recognize complex geometric patterns.

[0121] The input to this module is a high-resolution subgraph and a low-resolution subgraph, denoted as [missing information]. The node features are Adjacency relationships and edge weights Depend on Definition. A two-layer GAT architecture is adopted, which integrates node features from... Gradually updated to Ultimately, the enhanced features are obtained. .

[0122] For any node p in the subgraph and its features at level l Its features down to the next layer The update is the core of this module. Standard graph convolution treats all neighbors equally, while graph attention introduces a learnable attention coefficient. This is used to quantify the importance of neighbor nodes q to the center node p. First, a linear transformation is performed on the source and target nodes, projecting them onto a new feature space. :

[0123] (8)

[0124] In the formula, l is the layer number of the current graph attention layer; , Let p and q be the new feature representations after linear projection of the l-th layer; It is the learnable shared weight matrix of the l-th layer.

[0125] Subsequently, the raw attention scores for node pairs (p,q) are calculated. This is achieved through a single-layer feedforward neural network. To achieve this, the network is spliced ​​together with the transformed features. As input, followed by a LeakyReLU nonlinear activation function:

[0126] (9)

[0127] In the formula, In the l-th layer of the graph attention mechanism, the original attention score of point p to its neighbor point q; , Let p and q be the new feature representations after linear projection of the l-th layer; For feature splicing operations; Let be the attention vector of the l-th layer; LeakyReLU is the ReLU activation function with leakage.

[0128] To facilitate comparison of attention coefficients between different nodes, the attention coefficients of node p are then compared in the neighbor set. Normalization is performed using the Softmax function:

[0129] (10)

[0130] In the formula, In the l-th layer of the graph attention mechanism, the normalized score of point p with respect to its neighbor q; Let p be the original attention score of point p to its neighbor q. Let p be the original attention score of point p to its neighbor point r; Let be the actual neighborhood set of point p.

[0131] To further enhance the model's perception of geometric structures, curvature perception maps will be used... Prior edge weights inherited from These edge weights encode the initial geometric similarity between nodes, providing a stable and physically meaningful bias for the attention mechanism. The final standardized attention coefficients are defined as follows:

[0132] (11)

[0133] In the formula, The attention weights ultimately used in the l-th layer of the graph attention mechanism; In the l-th layer of the graph attention mechanism, the normalized score of point p with respect to its neighbor q; Let be the final edge weight between point p and point q.

[0134] Finally, the feature of node p at layer l+1 is the weighted sum of the transformed features of all its neighboring nodes, followed by a nonlinear activation function σ:

[0135] (12)

[0136] In the formula, Let σ represent the new node features of point p in the (l+1)th layer (i.e., after the update); σ represents the nonlinear activation function. The attention weights ultimately used in the l-th layer of the graph attention mechanism; The new feature representation of point p after linear projection through the l-th layer; Let be the actual neighborhood set of point p.

[0137] To stabilize the learning process and capture relevant information in different subspaces, multi-head attention is typically employed. Specifically, the attention transformation described above is performed K times independently, and then all the resulting feature vectors are concatenated (or averaged) to form the final output feature.

[0138] (13)

[0139] In the formula, Let K be the final node feature of point p in the (l+1)th layer; K is the number of multi-head attention heads. This is the feature concatenation operation; k is the current head number; σ represents the non-linear activation function; The attention weights ultimately used by the k-th head in layer l; It is the weight matrix of the k-th head in the l-th layer; The input features are for the residence q in the l-th layer; Let be the actual neighborhood set of point p.

[0140] In the architecture, two layers of GAT each have their own function: the first layer (l=0) mainly focuses on the uniformity of local geometric features, such as aggregating neighboring high-curvature points to enhance edge responses; the second layer (l=1) expands the receptive field, aiming to capture more global patterns. When faced with high-intensity noise, the attention mechanism can adaptively shrink the weights to at least a few geometrically or semantically highly reliable neighbors, thus exhibiting excellent robustness. After iterative propagation through the two layers of GAT, for high-resolution subgraphs, node features from the initial... Updated to Focusing on local details (such as millimeter-level edges). For low-resolution subgraphs, node features are derived from the initial... Updated to It captures global patterns (such as tunnel background curvature). It carries richer and more structured contextual information, providing a high-quality representation foundation for the final query fusion module.

[0141] For the GeoAdaptive Transformer Decoder, it inherits the refined multi-resolution sub-graph features output by the GAT enhancement module, i.e., high-resolution enhanced features. and low-resolution enhancement features The Geometric Adaptive Transformer Decoder (GAT-Decoder) plays a crucial role in generating the final segmentation prediction. This decoder is specifically designed to address two core challenges in the segmentation of small objects in point clouds: first, accurately focusing on tiny target regions against a broad background; and second, effectively fusing local geometric features extracted by ACQ-Net with global semantic information provided by U-Net.

[0142] The core innovation of GAT-Decoder lies in its geometrically adaptive attention mechanism. Unlike traditional Transformer decoders, this mechanism does not process relations at all locations equally. Instead, it dynamically allocates computational resources to regions that are more critical to the segmentation task by introducing geometric prior knowledge. Specifically, the decoder receives two key inputs from the GAT enhancement module: high-resolution query features... (Depend on (generated via linear projection) and low-resolution query features (Depend on (Generated via linear projection), and global features from the U-Net backbone network. The decoder achieves feature complementarity through a cross-resolution attention mechanism. High-resolution features provide detailed local geometric information, while low-resolution features contribute to global contextual understanding. The two are fused using learnable weights.

[0143] (14)

[0144] In the formula, The query features that are ultimately sent to the Transformer decoder; High-resolution query features; α represents the low-resolution query features; UpSample() is the nearest neighbor upsampling operation; α is the adaptive weight dynamically calculated based on the local features of the current point.

[0145] Then, standard attention calculations are performed:

[0146] (15)

[0147] In the formula, Basic attention output; These are query features; and It is based on global features Keys and values ​​obtained through linear transformation; It is the scaling factor; Softmax() is the activation function.

[0148] However, the key innovation lies in the introduction of a geometry-aware masking mechanism. This mask is based on curvature features computed in ACQ-Net. This generation process allows high-curvature regions 'a' (typically where edges or small objects are located) to receive higher attention weights.

[0149] (16)

[0150] In the formula, This is the final geometric mask matrix; and These are learnable parameters; It is the curvature eigenvector; Sigmoid() is a function that can compress any real number into the interval (0,1).

[0151] This geometric mask is then element-wise multiplied with the base attention weights to achieve geometric guidance of the attention distribution:

[0152] (17)

[0153] In the formula, This is the final geometrically adaptive attention matrix used for feature aggregation; Basic attention output; It is a geometric mask matrix; This indicates element-wise multiplication.

[0154] The fused adaptive attention output It is directly used as the core result of this layer and inherits from it. Multi-scale semantic content, simultaneously Dynamic scaling: This amplifies output intensity at high curvature points (enhancing edge response) and suppresses noise in low curvature regions, thereby improving local segmentation accuracy and ensuring seamless integration of geometric priors into the fused multi-resolution contextual representation. Subsequently, It is input into a feedforward network (FFN) for further nonlinear transformation, serving as the basis for the output of this layer.

[0155] The decoder employs a three-layer stacked structure, with each layer containing geometrically adaptive attention, a feedforward neural network, and residual connections. The feedforward network uses a standard two-layer MLP architecture.

[0156] (18)

[0157] In the formula, The features are the new features after passing through the feedforward network; ReLU is the activation function. and It is the learnable weight matrix and bias of the first layer MLP. and is the learnable weight matrix and bias of the second-layer MLP; x is the attention output of the previous layer after geometric masking.

[0158] Residual connections and layer normalization ensure training stability:

[0159] (19)

[0160] In the formula, These are the input features of this layer; Indicates sub-level operations; This represents the output feature of the layer; LayerNorm is the layer normalization operation. This mechanism alleviates the gradient vanishing problem through residual skipping and normalization, promoting stable convergence of deep networks.

[0161] After three layers of decoding (each layer sequentially applies geometrically adaptive attention generation) (FFN transformation and accumulation through residual connections) The model outputs a segmentation prediction for each point, and the prediction results are restored to the original point cloud resolution through upsampling operation to generate the probability of each point belonging to the target category.

[0162] For the loss function design, a multi-task loss function specifically designed for the segmentation of small targets in point clouds was developed to effectively train the entire segmentation network. This loss function not only considers the overall segmentation accuracy but also pays special attention to the boundary quality of small targets and the feature learning process.

[0163] The segmentation loss uses a weighted cross-entropy approach to address the extreme imbalance between foreground and background points in point cloud data.

[0164] (20)

[0165] in, The segmentation loss is the weighted cross-entropy; N is the total number of points in the current batch that participate in the loss calculation. It is the true label of the i-th point; It is the network's predicted probability for the i-th point; and These are the weights of the positive and negative samples, respectively.

[0166] Boundary augmentation loss is specifically designed to improve the segmentation quality of small object boundaries. This loss term imposes stronger constraints in high-curvature regions:

[0167] (twenty one)

[0168] In the formula, Increase the loss at the boundary; It is a set of high curvature points, passed through a curvature threshold. Sure; It is the true label of the i-th point; This is the network's predicted probability for the i-th point. This design ensures that the model produces sharper and more accurate segmentation results at the target boundary.

[0169] In addition, a dedicated supervision loss was designed for the sampling process of ACQ-Net:

[0170] (twenty two)

[0171] In the formula, is the sampling supervision loss for ACQ-Net; M is the number of points currently participating in training; These are the sample scores generated by ACQ-Net; These are the true labels of the corresponding points; the variance term Used to encourage diversity in sampling scores; λ is a hyperparameter that controls the strength of variance regularization, preventing the model from always selecting the same type of points.

[0172] The total loss function is a weighted combination of the losses mentioned above:

[0173] (twenty three)

[0174] In the formula, This represents the final total loss; The weighted cross-entropy segmentation loss; Increase the loss at the boundary; The sampling supervision loss for ACQ-Net; , These are the boundary loss weights and the sampling loss weights, respectively.

[0175] This multi-task loss design ensures that the model maintains high overall segmentation accuracy while significantly improving the ability to identify small targets. It also optimizes the sampling effect of ACQ-Net, thus guaranteeing the excellent performance of the entire system in the task of segmenting small targets in tunnel point clouds.

[0176] In other words, this invention addresses the challenges of sparse distribution, noise interference, and multi-scale fusion in small target segmentation of tunnel point clouds. It systematically constructs a single-stream segmentation network framework, achieving efficient processing from raw data to refined prediction. The sparse 3D U-Net backbone extracts compressed global semantic representations through symmetric encoding-decoding paths and skip connections, effectively alleviating the burden of large-scale computation. The graph structure construction and two-stage sampling mechanism of the adaptive curvature query network cleverly bridge task-oriented filtering and geometric dependency inheritance, ensuring that key relationships are fully preserved in high- and low-resolution subgraphs. The two-layer multi-head mechanism of the graph attention enhancement module iteratively improves the robustness and structured representation of node features through prior edge weight bias and adaptive aggregation. The three-layer stacked design of the geometrically adaptive Transformer decoder uses fusion query-driven attention computation and applies curvature mask post-processing to dynamically enhance the response in high-curvature regions, while maintaining deep stability through residual connections and layer normalization. The nonlinear extension of the feedforward network further amplifies the feature capacity and avoids information degradation. The combined strategy of multi-task loss functions not only balances overall accuracy and boundary sharpness but also optimizes the diversity of the sampling process, avoiding mode collapse. The framework demonstrates excellent generalization ability and practical applicability in benchmark validation, verifying the synergistic potential of sparse convolution, geometrically perceptual maps, and the Transformer paradigm, providing a scalable paradigm for similar sparse 3D data tasks. Looking ahead, further exploration of lightweight deployment strategies and multimodal data integration can drive the evolution of underground engineering inspection towards real-time intelligence.

[0177] Implementation Cases

[0178] This invention has been successfully implemented and is operating stably for a long time in an in-service highway tunnel in Sichuan. The tunnel uses a vehicle-mounted 3D laser scanning system to collect approximately 82 million point clouds in a single scan, including the lining, pipeline supports, drainage ditches, lighting cable trays, and multiple cracks and seepage defects. The system is deployed directly with the vehicle-mounted equipment, and the entire segmentation process is completed end-to-end without manual intervention. This invention's solution was compared with a traditional manual + semi-automatic inspection solution over a six-month parallel comparative verification process. The raw point clouds collected by the same vehicle-mounted laser scanning equipment were processed by both systems simultaneously. The results are shown in the table below:

[0179] Table 1 Comparison between the present invention solution and the original inspection solution

[0180] project Existing inspection plan Invention Solution Actual improvement Single full tunnel processing time 2-3 days 20 minutes + half a day for manual review Time reduced by more than 80% Bolt integrity detection rate 98.4% 97.7% -0.7% ≥0.8 mm crack continuous tracking rate 87.3% 91.1% +3.8% Seepage area positioning error Average 0.6 m Average 0.39 m Error reduced by 35% Discover the number of subtle defects that are easily missed by manual methods 11 places 38 places 27 new sites discovered

[0181] After voxelization of the original point cloud, it is input into a sparse 3D U-Net. Convolution calculations are performed only on approximately 1.3 million non-empty voxels. On a single RTX 4090 GPU, the peak memory usage is approximately 16 GB, and the forward inference time for the entire point cloud is approximately 20 minutes. The system first accurately calculates curvature based on the neighborhood normal vector deviation and constructs a curvature-aware map, so that the bolt edges, crack sides, and metal component outlines naturally form strongly connected subgraphs in the graph structure. Subsequently, under the constraint of a dedicated sampling loss, ACQ-Net increases the proportion of real foreground points in the high-resolution subgraph (8192 points) to approximately 62%, basically covering all anchor bolt outlines and major crack segments. A two-layer multi-head graph attention network with geometric edge weights further purifies the features of high-curvature regions. A three-layer curvature mask-guided Transformer decoder, combined with a point-by-point adaptive query fusion mechanism, ensures that the bolt and pipeline support outlines are complete, the continuity of cracks larger than 0.8 mm is good, the boundaries of seepage areas are clear, and there are almost no obvious missegmented patches on the lining surface.

[0182] In actual operation, there are three most convincing examples:

[0183] 1. A 0.7 mm × 2.8 m circumferential microcrack in the tunnel arch was not recorded in three consecutive inspections using traditional methods, but was completely tracked and marked in the first inspection by this system.

[0184] 2. At K124+680, there are 6 anchor bolts that are covered by water seepage and are almost invisible to the naked eye. Traditional methods can only find 3 of them, but this system can identify all 6 at once and their outlines are complete.

[0185] 3. A shallow seepage area of ​​about 0.9 m² on the left side wall was ignored in the traditional solution because its color was similar to the lining. This system stably divided the area and tracked the area change for 3 consecutive months.

[0186] The system ultimately outputs a complete point cloud and a two-dimensional defect distribution map with semantic colors, and has been the core tool for daily inspections of the tunnel for several months. Compared to the previous method of relying mainly on manual visual inspection and spot checks, the system now only requires inspectors to focus on reviewing the suspicious areas marked by the system. The time for a single full tunnel inspection has been reduced from 2-3 days to less than half a day. At the same time, the system has repeatedly captured minute circumferential cracks, local water seepage, and minor lining spalling that are easily missed by manual inspections. The detection rate and efficiency have been significantly improved. The generated defect distribution map has been directly included in the monthly safety assessment report and has been unanimously recognized by the owner and operation and maintenance unit. It is currently operating stably in the tunnel.

[0187] To achieve the above objectives, the present invention also provides a tunnel point cloud small target segmentation processing system based on a curvature-aware adaptive network, such as... Figure 4 As shown, the system is applied to the tunnel point cloud small target segmentation processing method based on curvature-aware adaptive network, and the system includes:

[0188] The data creation and processing unit is used to create and acquire first data corresponding to the tunnel, and process the first data through a sparse 3D U-Net backbone network voxelization, while generating second data corresponding to the first data; wherein, the first data is tunnel point cloud data; the second data includes multi-scale global semantic feature data and three-dimensional coordinate feature data;

[0189] A data construction and processing unit is used to construct a curvature-aware map corresponding to the second data, and to fuse and process the second data and the third data generated by calculating the deviation of the neighborhood normal vector; wherein, the third data is curvature feature data;

[0190] The first data processing and generation unit is used to combine the adaptive curvature query network with differentiable sampling to process the curvature-aware map and generate fourth data corresponding to the curvature-aware map; wherein, the fourth data is a high-resolution sub-map and a low-resolution sub-map;

[0191] The second data processing and generation unit is used to enhance the fourth data through graph attention network features, generate fifth data corresponding to the fourth data, and based on the fifth data and the second data, create and generate a sixth data corresponding to the tunnel through a geometric adaptive attention mechanism and curvature mask guidance; wherein, the fifth data is enhanced multi-resolution feature data; and the sixth data is segmentation prediction result data.

[0192] Furthermore, the sparse 3D U-Net backbone network is processed using sparse convolution operations and computation is performed only on non-empty voxels;

[0193] The sparse 3D U-Net backbone network has an encoder-decoder architecture with skip connections; wherein the encoder and the decoder are symmetrically arranged, the encoder achieves progressive downsampling through multiple layers of sparse convolution, and the decoder achieves progressive upsampling through transposed convolution.

[0194] Each layer of the decoder is fused with the output features of the corresponding layer of the encoder via skip connections;

[0195] The data construction and processing unit further includes:

[0196] The first generation module is used to generate curvature features corresponding to each node by summing and normalizing the normal vector projection deviations within the k-nearest neighbor range; wherein, the node is each non-empty voxel output by the sparse 3D U-Net;

[0197] The first processing module is used to construct a graph topology structure corresponding to the curvature-aware graph based on the initial node features that integrate global semantic features and the curvature features, using the k-nearest neighbor algorithm, wherein the weight value of the edge is determined by the spatial proximity and geometric consistency between nodes.

[0198] And / or, the first data processing and generation unit further includes:

[0199] The second generation module is used to combine a multilayer perceptron with shared weights to create seventh data corresponding to each node in the curvature perception map; wherein, the seventh data is an importance score.

[0200] The third generation module is used to convert the seventh data into a normalized sampling probability distribution using a differentiable method, and based on the sampling probability distribution, combined with hierarchical filtering processing, to create a node index set corresponding to the probability nodes; wherein, the node index set includes a high-resolution node index set and a low-resolution node index set.

[0201] The fourth generation module is used to induce the generation of fourth data corresponding to the curvature-aware map based on the node index set;

[0202] And / or, the second data processing generation unit further includes:

[0203] The second processing module is used to create and generate eighth data corresponding to the high-resolution subgraph or the low-resolution subgraph respectively, and to adaptively weight-fuse the eighth data; wherein, the eighth data includes high-resolution query features and low-resolution query features;

[0204] The third processing module is used to generate a corresponding geometric perception mask based on the third data, and to perform element-wise multiplication of the mask with the basic attention weights generated through query-key interaction; thereby realizing geometric prior guidance of attention distribution.

[0205] The fourth processing module is used to input the enhanced multi-resolution features and the global features extracted by the sparse 3D U-Net backbone network into the geometrically adaptive Transformer decoder; wherein, the geometrically adaptive Transformer decoder adopts a three-layer stacked structure, each layer containing a geometrically adaptive attention module, a feedforward network and a residual connection; the feedforward network adopts a two-layer MLP architecture;

[0206] The graph attention network employs a multi-head graph attention layer, incorporating edge weights inherited from the curvature-aware graph as attention biases during feature propagation, and updating node features through two iterative layers.

[0207] In the system solution embodiment of the present invention, the specific details of the method steps involved in the tunnel point cloud small target segmentation processing based on curvature-aware adaptive network have been described above. That is to say, the functional modules in the system are used to implement the steps or sub-steps in the above method embodiment, and will not be repeated here.

[0208] To achieve the above objectives, the present invention also provides a tunnel point cloud small target segmentation processing platform based on a curvature-aware adaptive network, such as... Figure 5 As shown, the system includes a processor, a memory, and a control program for a tunnel point cloud small target segmentation processing platform based on a curvature-aware adaptive network. The processor executes the control program, which is stored in the memory. This control program implements the steps of the tunnel point cloud small target segmentation processing method based on a curvature-aware adaptive network, for example:

[0209] S01. Create and acquire first data corresponding to the tunnel, and process the first data by voxelization through a sparse 3D U-Net backbone network, while generating second data corresponding to the first data; wherein, the first data is tunnel point cloud data; the second data includes multi-scale global semantic feature data and three-dimensional coordinate feature data;

[0210] S02. Construct a curvature-aware map corresponding to the second data, and fuse the second data and the third data generated by calculating the deviation of the neighborhood normal vector; wherein, the third data is curvature feature data;

[0211] S03. The curvature-aware map is processed by differentiable sampling using an adaptive curvature query network, and fourth data corresponding to the curvature-aware map is generated; wherein, the fourth data consists of a high-resolution sub-map and a low-resolution sub-map.

[0212] S04. The fourth data is enhanced by graph attention network feature enhancement to generate the fifth data corresponding to the fourth data. Based on the fifth data and the second data, a sixth data corresponding to the tunnel is created and generated through geometric adaptive attention mechanism and curvature mask guidance. The fifth data is the enhanced multi-resolution feature data. The sixth data is the segmentation prediction result data.

[0213] The specific details of the steps have been explained above and will not be repeated here.

[0214] In this embodiment of the invention, the built-in processor of the tunnel point cloud small target segmentation processing platform based on curvature-aware adaptive networks can be composed of integrated circuits. For example, it can be composed of a single packaged integrated circuit, or multiple integrated circuits packaged with the same or different functions. This includes combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor connects to various components using various interfaces and lines, and executes programs or units stored in memory, as well as calling data stored in memory, to perform various functions and process data for tunnel point cloud small target segmentation based on curvature-aware adaptive networks.

[0215] The memory is used to store program code and various data. It is installed in the tunnel point cloud small target segmentation processing platform based on curvature-aware adaptive networks and enables high-speed and automatic access to programs or data during operation. The memory includes read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically erasable programmable read-only memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0216] To achieve the above objectives, the present invention also provides a computer-readable storage medium, such as... Figure 6 As shown, the computer-readable storage medium stores a control program for a tunnel point cloud small target segmentation processing platform based on a curvature-aware adaptive network. This control program implements the steps of the tunnel point cloud small target segmentation processing method based on a curvature-aware adaptive network; for example:

[0217] S01. Create and acquire first data corresponding to the tunnel, and process the first data by voxelization through a sparse 3D U-Net backbone network, while generating second data corresponding to the first data; wherein, the first data is tunnel point cloud data; the second data includes multi-scale global semantic feature data and three-dimensional coordinate feature data;

[0218] S02. Construct a curvature-aware map corresponding to the second data, and fuse the second data and the third data generated by calculating the deviation of the neighborhood normal vector; wherein, the third data is curvature feature data;

[0219] S03. The curvature-aware map is processed by differentiable sampling using an adaptive curvature query network, and fourth data corresponding to the curvature-aware map is generated; wherein, the fourth data consists of a high-resolution sub-map and a low-resolution sub-map.

[0220] S04. The fourth data is enhanced by graph attention network feature enhancement to generate the fifth data corresponding to the fourth data. Based on the fifth data and the second data, a sixth data corresponding to the tunnel is created and generated through geometric adaptive attention mechanism and curvature mask guidance. The fifth data is the enhanced multi-resolution feature data. The sixth data is the segmentation prediction result data.

[0221] The specific details of the steps have been explained above and will not be repeated here.

[0222] In the description of embodiments of the present invention, it should be noted that any process or method description in the flowcharts or otherwise described herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the present invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order according to the functions involved, as should be understood by those skilled in the art to which the embodiments of the present invention pertain.

[0223] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processing module, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, a “computer-readable medium” can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, the computer-readable medium can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0224] This invention creates and acquires first data corresponding to a tunnel through a method, processes the first data using a sparse 3D U-Net backbone network for voxelization, and simultaneously generates second data corresponding to the first data. The first data is tunnel point cloud data; the second data includes multi-scale global semantic feature data and three-dimensional coordinate feature data. A curvature-aware map corresponding to the second data is constructed, and the second data and third data generated by calculating the deviation of neighborhood normal vectors are fused together. The third data is curvature feature data. The curvature-aware map is processed by differentiable sampling using an adaptive curvature query network, and a fourth data corresponding to the curvature-aware map is generated. The fourth data consists of a high-resolution sub-map and a low-resolution sub-map. The fourth data is enhanced by a graph attention network to generate a fifth data corresponding to the fourth data. Based on the fifth data and the second data, a sixth data corresponding to the tunnel is created and generated using a geometric adaptive attention mechanism and curvature mask guidance. The fifth data is enhanced multi-resolution feature data; the sixth data is segmentation prediction result data; and the method includes a corresponding system, platform, and storage medium to significantly improve the accuracy and robustness of small target segmentation.

[0225] In other words, this invention constructs an end-to-end point cloud segmentation framework by introducing a sparse 3D U-Net backbone network, a curvature-aware graph structure, an adaptive differentiable sampling mechanism, and a geometry-guided Transformer decoder. This framework can efficiently and accurately identify and segment small targets such as bolt holes and cracks in massive, sparse, and noisy tunnel point clouds, improving the accuracy, robustness, and computational efficiency of small target segmentation in complex engineering environments, and providing reliable technical support for intelligent tunnel inspection.

[0226] In other words, against the backdrop of rapid urbanization and the digital transformation of infrastructure, the safety monitoring and maintenance of tunnel engineering has become an urgent need to ensure the stable operation of the transportation system. The internal environment of tunnels is complex and variable, susceptible to geological stress, construction errors, and long-term corrosion, causing minute defects such as bolt holes and cracks to be hidden within the vast concrete background. Failure to identify these defects in a timely and accurate manner can lead to serious safety hazards. Point cloud data generated by lidar, with its millimeter-level accuracy and non-contact acquisition advantages, has become the preferred method for tunnel defect detection. However, this data is often massive, sparsely distributed, and densely noisy, with far more background points than foreground targets. Traditional segmentation methods, such as point cloud-based deep networks or dense convolutional models, struggle to balance efficiency and accuracy: the former tends to ignore nonlocal geometric relationships, while the latter falls into the trap of high computational overhead and cannot effectively focus on the edge features of small targets in noisy environments. To address these challenges, this invention proposes an innovative single-stream segmentation network framework, using a sparse 3D U-Net as a high-efficiency backbone. It operates only on non-empty voxels, significantly compressing redundancy and extracting multi-scale global semantic features, while incorporating geometric injection of three-dimensional coordinates to enhance spatial perception. The upstream module, the Adaptive Curvature Query Network, dynamically filters high-value anchor point subgraphs through curvature-aware map construction and a differentiable sampling mechanism, achieving hierarchical capture of local details and macroscopic patterns. The downstream module, the Geometric Adaptive Transformer Decoder, introduces curvature mask-guided attention allocation, integrates multi-resolution queries and U-Net global representations, and progressively refines prediction boundaries through a three-layer stacked structure. The framework's multi-task loss design further enhances the optimization of class imbalance and sampling diversity, ensuring sharp and robust segmentation output in complex tunnel scenarios. Experimental results demonstrate that this method exhibits superior segmentation accuracy and computational efficiency compared to benchmarks on real-world datasets, providing a solid technical foundation for the engineering deployment of intelligent tunnel inspection systems.

[0227] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for segmenting small targets in tunnel point clouds based on curvature-aware adaptive networks, characterized in that, The method includes: First data corresponding to the tunnel is created and acquired, and the first data is processed by voxelization through a sparse 3D U-Net backbone network. At the same time, second data corresponding to the first data is generated. The first data is tunnel point cloud data. The second data includes multi-scale global semantic feature data and three-dimensional coordinate feature data. A curvature-aware map corresponding to the second data is constructed, and the second data and the third data generated by calculating the deviation of the neighborhood normal vector are fused and processed; wherein, the third data is curvature feature data; The curvature-aware map is processed by a differentiable sampling method using an adaptive curvature query network, and fourth data corresponding to the curvature-aware map is generated; wherein, the fourth data consists of a high-resolution sub-map and a low-resolution sub-map. The fourth data is enhanced by graph attention network feature enhancement to generate a fifth data corresponding to the fourth data. Based on the fifth data and the second data, a sixth data corresponding to the tunnel is created and generated through geometric adaptive attention mechanism and curvature mask guidance. The fifth data is enhanced multi-resolution feature data, and the sixth data is segmentation prediction result data.

2. The method for segmenting small targets in tunnel point clouds based on curvature-aware adaptive networks according to claim 1, characterized in that, The sparse 3D U-Net backbone network is processed using sparse convolution operations and only performs computations on non-empty voxels. The sparse 3D U-Net backbone network has an encoder-decoder architecture with skip connections; wherein the encoder and the decoder are symmetrically arranged, the encoder achieves progressive downsampling through multiple layers of sparse convolution, and the decoder achieves progressive upsampling through transposed convolution. Each layer of the decoder is fused with the output features of the corresponding layer of the encoder via skip connections.

3. A method for segmenting small targets in tunnel point clouds based on a curvature-aware adaptive network according to claim 1 or 2, characterized in that, The construction of the curvature-aware map corresponding to the second data, and the fusion processing of the second data and the third data generated by calculating the deviation of the neighborhood normal vector, further includes: Curvature features are generated by summing and normalizing the normal vector projection deviations within the k-nearest neighbor range, and are corresponding to each node; wherein, the node is each non-empty voxel output by the sparse 3D U-Net. Based on the initial node features that integrate global semantic features and curvature features, a graph topology corresponding to the curvature-aware graph is constructed using the k-nearest neighbor algorithm, wherein the weight values ​​of the edges are determined by the spatial proximity and geometric consistency between nodes.

4. The method for segmenting small targets in tunnel point clouds based on curvature-aware adaptive networks according to claim 1, characterized in that, The method of combining the adaptive curvature query network with differentiable sampling to process the curvature-aware map and generate fourth data corresponding to the curvature-aware map further includes: By combining a multilayer perceptron with shared weights, a seventh data point corresponding to each node in the curvature-sensing graph is created; wherein, the seventh data point is an importance score. The seventh data is converted into a normalized sampling probability distribution using a differentiable method. Based on the sampling probability distribution and combined with hierarchical filtering, a node index set corresponding to the probability nodes is created. The node index set includes a high-resolution node index set and a low-resolution node index set. Based on the node index set, fourth data corresponding to the curvature-sensing map is induced to be generated.

5. The method for segmenting small targets in tunnel point clouds based on curvature-aware adaptive networks according to claim 1, characterized in that, The process of enhancing the fourth data through graph attention network features to generate fifth data corresponding to the fourth data, and based on the fifth data and the second data, creating and generating sixth data corresponding to the tunnel through a geometric adaptive attention mechanism and curvature mask guidance, further includes: Eighth data corresponding to either the high-resolution subgraph or the low-resolution subgraph is created and generated, and the eighth data is processed by adaptive weight fusion; wherein, the eighth data includes both high-resolution query features and low-resolution query features; Based on the third data, a corresponding geometric perception mask is generated, and the mask is multiplied element-wise with the basic attention weights generated through query-key interaction.

6. A method for segmenting small targets in tunnel point clouds based on a curvature-aware adaptive network according to claim 1 or 5, characterized in that, The process of enhancing the fourth data through graph attention network features to generate fifth data corresponding to the fourth data, and based on the fifth data and the second data, creating and generating sixth data corresponding to the tunnel through a geometric adaptive attention mechanism and curvature mask guidance, further includes: The enhanced multi-resolution features and the global features extracted by the sparse 3D U-Net backbone network are input into the geometrically adaptive Transformer decoder; wherein, the geometrically adaptive Transformer decoder adopts a three-layer stacked structure, each layer containing a geometrically adaptive attention module, a feedforward network and a residual connection; the feedforward network adopts a two-layer MLP architecture; The graph attention network employs a multi-head graph attention layer, incorporating edge weights inherited from the curvature-aware graph as attention biases during feature propagation, and updating node features through two iterative layers.

7. A tunnel point cloud small target segmentation processing system based on curvature-aware adaptive network, characterized in that, The system is applied to the tunnel point cloud small target segmentation processing method based on curvature-aware adaptive network as described in any one of claims 1 to 6, and the system comprises: The data creation and processing unit is used to create and acquire first data corresponding to the tunnel, and process the first data through a sparse 3D U-Net backbone network voxelization, while generating second data corresponding to the first data; wherein, the first data is tunnel point cloud data; the second data includes multi-scale global semantic feature data and three-dimensional coordinate feature data; A data construction and processing unit is used to construct a curvature-aware map corresponding to the second data, and to fuse and process the second data and the third data generated by calculating the deviation of the neighborhood normal vector; wherein, the third data is curvature feature data; The first data processing and generation unit is used to combine the adaptive curvature query network with differentiable sampling to process the curvature-aware map and generate fourth data corresponding to the curvature-aware map; wherein, the fourth data is a high-resolution sub-map and a low-resolution sub-map; The second data processing and generation unit is used to enhance the fourth data through graph attention network features, generate fifth data corresponding to the fourth data, and based on the fifth data and the second data, create and generate a sixth data corresponding to the tunnel through a geometric adaptive attention mechanism and curvature mask guidance; wherein, the fifth data is enhanced multi-resolution feature data; and the sixth data is segmentation prediction result data.

8. The tunnel point cloud small target segmentation processing system based on curvature-aware adaptive network according to claim 7, characterized in that, The sparse 3D U-Net backbone network is processed using sparse convolution operations and only performs computations on non-empty voxels. The sparse 3D U-Net backbone network has an encoder-decoder architecture with skip connections; wherein the encoder and the decoder are symmetrically arranged, the encoder achieves progressive downsampling through multiple layers of sparse convolution, and the decoder achieves progressive upsampling through transposed convolution. Each layer of the decoder is fused with the output features of the corresponding layer of the encoder via skip connections; The data construction and processing unit further includes: The first generation module is used to generate curvature features corresponding to each node by summing and normalizing the normal vector projection deviations within the k-nearest neighbor range; wherein, the node is each non-empty voxel output by the sparse 3D U-Net; The first processing module is used to construct a graph topology structure corresponding to the curvature-aware graph based on the initial node features that integrate global semantic features and the curvature features, using the k-nearest neighbor algorithm, wherein the weight value of the edge is determined by the spatial proximity and geometric consistency between nodes. And / or, the first data processing and generation unit further includes: The second generation module is used to combine a multilayer perceptron with shared weights to create seventh data corresponding to each node in the curvature perception map; wherein, the seventh data is an importance score. The third generation module is used to convert the seventh data into a normalized sampling probability distribution using a differentiable method, and based on the sampling probability distribution, combined with hierarchical filtering processing, to create a node index set corresponding to the probability nodes; wherein, the node index set includes a high-resolution node index set and a low-resolution node index set. The fourth generation module is used to induce the generation of fourth data corresponding to the curvature-aware map based on the node index set; And / or, the second data processing generation unit further includes: The second processing module is used to create and generate eighth data corresponding to the high-resolution subgraph or the low-resolution subgraph respectively, and to adaptively weight-fuse the eighth data; wherein, the eighth data includes high-resolution query features and low-resolution query features; The third processing module is used to generate a corresponding geometric perception mask based on the third data, and to perform element-wise multiplication of the mask with the basic attention weights generated by query-key interaction. The fourth processing module is used to input the enhanced multi-resolution features and the global features extracted by the sparse 3D U-Net backbone network into the geometrically adaptive Transformer decoder; wherein, the geometrically adaptive Transformer decoder adopts a three-layer stacked structure, each layer containing a geometrically adaptive attention module, a feedforward network and a residual connection; the feedforward network adopts a two-layer MLP architecture; The graph attention network employs a multi-head graph attention layer, incorporating edge weights inherited from the curvature-aware graph as attention biases during feature propagation, and updating node features through two iterative layers.

9. A tunnel point cloud small target segmentation processing platform based on curvature-aware adaptive network, characterized in that, The system includes a processor, a memory, and a control program for a tunnel point cloud small target segmentation processing platform based on a curvature-aware adaptive network. The processor executes the control program, which is stored in the memory. The control program implements the tunnel point cloud small target segmentation processing method based on a curvature-aware adaptive network as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a control program for a tunnel point cloud small target segmentation processing platform based on a curvature-aware adaptive network. The control program for the tunnel point cloud small target segmentation processing platform based on a curvature-aware adaptive network implements the tunnel point cloud small target segmentation processing method based on a curvature-aware adaptive network as described in any one of claims 1 to 6.

Images