SPDCN: Strip-based Deformable Convolutional Network for Steel Surface Defect Segmentation

Researchers introduced SPDCN, a compact steel surface defect segmentation model designed to better detect thin, irregular flaws.

The model combines a fuzzy-enhanced multi-scale context module with adaptive direction-aware deformable convolution for anisotropic defects.

SPDCN reported state-of-the-art performance on public benchmarks, including 89.60% mIoU on NEU-Seg with just 3.54M parameters.

The approach could improve industrial inspection by catching cracks, scratches, and other defects that standard convolutions often miss.