DMMCNet: dynamic multiscale and multilevel contrastive point cloud feature extraction network for industrial defect detection

Industrial defect detection is crucial for ensuring product quality and safety. Point-cloud-based methods have attracted increasing attention due to their accurate geometric feature extraction capability. However, most of them (1) lack the ability to dynamically adjust multi-scale receptive fields,...

Descripción completa

Detalles Bibliográficos
Autores: Chen, Wenhui, Liu, Jirui, Ji, Zhenyan, Song, Yuezeng, Armendáriz Íñigo, José Enrique, Liu, Jiqiang
Tipo de recurso: artículo
Estado:Versión aceptada para publicación
Fecha de publicación:2026
País:España
Institución:Universidad San Jorge (USJ)
Repositorio:Academica-e. Repositorio Institucional de la Universidad Pública de Navarra
OAI Identifier:oai:dnet:academicae__::7c0820d11f46a32a77484942e3c5841f
Acceso en línea:https://hdl.handle.net/2454/56988
Access Level:acceso abierto
Palabra clave:Contrastive learning
Industrial defect detection
Multiscale feature learning
Point cloud
Descripción
Sumario:Industrial defect detection is crucial for ensuring product quality and safety. Point-cloud-based methods have attracted increasing attention due to their accurate geometric feature extraction capability. However, most of them (1) lack the ability to dynamically adjust multi-scale receptive fields, limiting their capacity to perceive defects with diverse sizes, (2) ignore semantic inconsistencies among multi-scale features during fusion, weakening the ability to distinguish defect regions, and (3) pay insufficient attention to key points, lacking sensitivity to fine-grained defects. To address these issues, we propose a Dynamic Multi-scale and Multi-level Contrastive Point Cloud Feature Extraction Network (DMMCNet) for industrial defect detection, which consists of three key modules: Dynamic Multi-scale Feature Encoding module (DMFE), Multi-level Contrastive Fusion module (MCF), and Key-Point Enhancement module (KPE). Specifically, in DMFE, we propose a learnable scale embedding mechanism that dynamically adjusts receptive field sizes according to the geometric context of the point cloud, improving the perception of defects with diverse sizes. In MCF, we develop a multi-level fusion scheme that progressively aggregates multi-scale features, allowing fine-scale defect cues to be preserved before incorporating background information. Within this fusion process, we propose a cross-scale contrastive constraint mechanism to align features from different receptive-field scales and alleviate semantic inconsistencies. In KPE, we model spatial correlations among points to guide the network's attention toward key points with distinct geometric variations, strengthening the response to critical defect areas. Finally, experimental results on three industrial point cloud datasets demonstrate that DMMCNet outperforms thirteen state-of-the-art methods.