CDARNet: A robust cross-dimensional adaptive region reconstruction network for real-time metal surface defect segmentation

Accurate and rapid detection of metal surface defects is crucial for ensuring product quality and improving production efficiency. Despite significant progress in vision-based detection methods, challenges such as excessive background noise, blurred boundaries, small defect sizes, interclass similarity, and intraclass difference continue to hinder effective defect recognition. Moreover, high-accuracy, real-time, and robust methods are urgently needed in industrial applications. To address these challenges, a robust cross-dimensional adaptive region reconstruction network (CDARNet) is proposed for real-time metal surface defect segmentation. CDARNet employs an encoder-decoder architecture. In the encoder, a cross-dimensional adaptive region reconstruction module is designed to mitigate interference and improve inference speed and robustness by adaptively reconstructing local regions across dimensions. It consists of the region partitioning and channel compression unit (RPCC) and the channel expansion and spatial reconstruction unit (CESR). RPCC suppresses irrelevant noise while preserving critical defect features by modeling intra-region relationships along the channel dimension. CESR expands the channel dimension to reconstruct spatial features and enrich defect representations. In the decoder, a contextual self-correlation semantic unification module is designed to enhance semantic correlation among pixels of the same defect category, improving localization and differentiation. Moreover, a cross-scale spatial feature refinement module refines boundaries and detects defects at various scales using both global and local features. Comprehensive experiments on multiple challenging datasets demonstrate that CDARNet not only achieves high detection accuracy with fewer parameters and superior robustness, but also exhibits competitive generalization across different defect types and complex environments, surpassing other leading methods.