GTAINet: Graph neural network-based two-stage anomaly identification for locking wire point clouds using hierarchical attentive edge convolution
With the application of 3D sensors, studies on various vision tasks based on point clouds have been explored in different fields. In the field of the high-speed train safety inspection, the vision-based artificial intelligence inspection technology has received greater attention in recent years. In...
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2022-12-01
|
| Series: | International Journal of Applied Earth Observations and Geoinformation |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1569843222002941 |
| _version_ | 1811190593597997056 |
|---|---|
| author | Zhixue Wang Yu Zhang Tongfei Lv Lin Luo |
| author_facet | Zhixue Wang Yu Zhang Tongfei Lv Lin Luo |
| author_sort | Zhixue Wang |
| collection | DOAJ |
| description | With the application of 3D sensors, studies on various vision tasks based on point clouds have been explored in different fields. In the field of the high-speed train safety inspection, the vision-based artificial intelligence inspection technology has received greater attention in recent years. In this paper, we introduce a graph neural network-based (GNN) two-stage anomaly identification method (GTAINet) for locking wire point clouds. GTAINet consists of two sub-networks, a segmentation network and a classification network. The segmentation network splits bolt, wire and background points, then the segmented bolt and wire points are fed into the classification network to identify whether the sample is broken. However, as an essential type of geometric data structure, how to extract the rich geometric representation from point clouds is the key to the above vision tasks. To this end, we proposed a hierarchical attentive edge convolution (HAEConv) to establish a GNN. HAEConv is able to recover hierarchical topological information from point clouds and attentively recalibrate each edge’s response, which allows the model to better capture more useful information of local and global geometric structure. Another critical point that limits the segmentation and classification performance is the lack of training data, in particular the lack of anomaly locking wire samples. To address this challenge, we propose a synthetic algorithm that can synthesize massive fake anomaly locking wire data using parameterized Bézier curves. Experiments demonstrate that the proposed networks based on HAEConv outperform popular existing methods on both segmentation and classification tasks. In addition, the synthetic method presented allows us to pre-train a model with very strong generalization ability, which can significantly improve model performance. |
| first_indexed | 2024-04-11T14:52:27Z |
| format | Article |
| id | doaj.art-0fbe3456025f458585b4a30770478c50 |
| institution | Directory Open Access Journal |
| issn | 1569-8432 |
| language | English |
| last_indexed | 2024-04-11T14:52:27Z |
| publishDate | 2022-12-01 |
| publisher | Elsevier |
| record_format | Article |
| series | International Journal of Applied Earth Observations and Geoinformation |
| spelling | doaj.art-0fbe3456025f458585b4a30770478c502022-12-22T04:17:21ZengElsevierInternational Journal of Applied Earth Observations and Geoinformation1569-84322022-12-01115103106GTAINet: Graph neural network-based two-stage anomaly identification for locking wire point clouds using hierarchical attentive edge convolutionZhixue Wang0Yu Zhang1Tongfei Lv2Lin Luo3School of Physical Science and Technology, Southwest Jiaotong University, Chengdu, ChinaCorresponding author.; School of Physical Science and Technology, Southwest Jiaotong University, Chengdu, ChinaSchool of Physical Science and Technology, Southwest Jiaotong University, Chengdu, ChinaSchool of Physical Science and Technology, Southwest Jiaotong University, Chengdu, ChinaWith the application of 3D sensors, studies on various vision tasks based on point clouds have been explored in different fields. In the field of the high-speed train safety inspection, the vision-based artificial intelligence inspection technology has received greater attention in recent years. In this paper, we introduce a graph neural network-based (GNN) two-stage anomaly identification method (GTAINet) for locking wire point clouds. GTAINet consists of two sub-networks, a segmentation network and a classification network. The segmentation network splits bolt, wire and background points, then the segmented bolt and wire points are fed into the classification network to identify whether the sample is broken. However, as an essential type of geometric data structure, how to extract the rich geometric representation from point clouds is the key to the above vision tasks. To this end, we proposed a hierarchical attentive edge convolution (HAEConv) to establish a GNN. HAEConv is able to recover hierarchical topological information from point clouds and attentively recalibrate each edge’s response, which allows the model to better capture more useful information of local and global geometric structure. Another critical point that limits the segmentation and classification performance is the lack of training data, in particular the lack of anomaly locking wire samples. To address this challenge, we propose a synthetic algorithm that can synthesize massive fake anomaly locking wire data using parameterized Bézier curves. Experiments demonstrate that the proposed networks based on HAEConv outperform popular existing methods on both segmentation and classification tasks. In addition, the synthetic method presented allows us to pre-train a model with very strong generalization ability, which can significantly improve model performance.http://www.sciencedirect.com/science/article/pii/S1569843222002941Anomaly identificationGraph neural networkPoint cloud generalizationHierarchical attentive edge convolutionData synthesis |
| spellingShingle | Zhixue Wang Yu Zhang Tongfei Lv Lin Luo GTAINet: Graph neural network-based two-stage anomaly identification for locking wire point clouds using hierarchical attentive edge convolution International Journal of Applied Earth Observations and Geoinformation Anomaly identification Graph neural network Point cloud generalization Hierarchical attentive edge convolution Data synthesis |
| title | GTAINet: Graph neural network-based two-stage anomaly identification for locking wire point clouds using hierarchical attentive edge convolution |
| title_full | GTAINet: Graph neural network-based two-stage anomaly identification for locking wire point clouds using hierarchical attentive edge convolution |
| title_fullStr | GTAINet: Graph neural network-based two-stage anomaly identification for locking wire point clouds using hierarchical attentive edge convolution |
| title_full_unstemmed | GTAINet: Graph neural network-based two-stage anomaly identification for locking wire point clouds using hierarchical attentive edge convolution |
| title_short | GTAINet: Graph neural network-based two-stage anomaly identification for locking wire point clouds using hierarchical attentive edge convolution |
| title_sort | gtainet graph neural network based two stage anomaly identification for locking wire point clouds using hierarchical attentive edge convolution |
| topic | Anomaly identification Graph neural network Point cloud generalization Hierarchical attentive edge convolution Data synthesis |
| url | http://www.sciencedirect.com/science/article/pii/S1569843222002941 |
| work_keys_str_mv | AT zhixuewang gtainetgraphneuralnetworkbasedtwostageanomalyidentificationforlockingwirepointcloudsusinghierarchicalattentiveedgeconvolution AT yuzhang gtainetgraphneuralnetworkbasedtwostageanomalyidentificationforlockingwirepointcloudsusinghierarchicalattentiveedgeconvolution AT tongfeilv gtainetgraphneuralnetworkbasedtwostageanomalyidentificationforlockingwirepointcloudsusinghierarchicalattentiveedgeconvolution AT linluo gtainetgraphneuralnetworkbasedtwostageanomalyidentificationforlockingwirepointcloudsusinghierarchicalattentiveedgeconvolution |