Sewer defect detection from 3D point clouds using a transformer-based deep learning model

Targeting the defect classification from 3D point clouds, this research develops a deep learning method named the Transformer-based point cloud classification network (TransPCNet) to obtain superior classification results. The developed TransPCNet primarily consists of the feature embedding module, the attention module, and the classification module, where the first two modules are to enhance the feature extraction and learning capability for assisting the classification module to classify the 3D point clouds more accurately. In addition, a novel loss function is proposed to support the TransPCNet by strengthening feature learning and tackling data imbalance. The effectiveness of the developed TransPCNet is demonstrated on a publicly available dataset with both real and synthetic point clouds. In comparison with other state-of-the-art methods, the TransPCNet outperforms others with improvements of over 13.6%, 15.2%, and 13.7% in terms of precision, recall, and F1-score on the overall dataset. Moreover, the TransPCNet is robust and efficient in different scenarios, where the synthetic data is beneficial to enhancing the detection accuracy on real datasets. Overall, this research contributes to developing TransPCNet to conduct 3D point cloud classification, resulting in a more accurate and effective result with great practical potential.